Model sources

For provider cookbooks, prompt formats, and research gaps by managed model, see the provider prompting guide tracker.

Weights reach mere.run in three local-first ways:

1. Managed pulls — cataloged Hugging Face snapshots installed into the local model store by mere.run model pull
2. Registered locations — read-only search roots or explicit canonical-ID bindings managed by mere.run model location
3. Local paths — a directory you point at yourself with --model, --model-root, or the command's equivalent option

There is no private model archive, no credentialed mirror, and no central host in between. Every managed model's source repository and pinned revision are in the catalog, so you can see exactly where each byte came from.

The canonical local model store is:

~/Library/Application Support/MereRun/models

Override that with MERERUN_MODELS_DIR or --models-root.

Registered locations augment the normal persisted primary store without copying payloads or creating symlinks. Search roots use <root>/<canonical-model-id>/ and require each model's managed manifest. Explicit bindings can point a canonical ID at an arbitrarily named directory; mere.run validates the checkpoint at registration time and stores its identity in ~/Library/Application Support/MereRun/model_locations.json without writing metadata into the external directory. Externally registered files remain read-only and are outside model remove, model gc, manifest repair, and storage-reclamation ownership.

Canonical managed model IDs

This is the authoritative public catalog list. It is kept in sync with ManagedModelCatalog.allSpecs, and the test suite fails if the table drifts from the runtime catalog used by mere.run model list, mere.run model capabilities --all, and mere.run model pull.

API-served catalog entries also own a typed profile for their task, serving engine, input/output modalities, tool and structured-output support, reasoning levels and mappings, context/output limits, and OpenAI compatibility behavior. /v1/models, request validation, agent eligibility, and the Pi provider all project from that profile. Runtime aliases and configured limits are applied as an effective overlay; they are not a second capability catalog.

Catalog category	Model ID
image	image-klein-nano
image	image-klein-max
image	image-klein-9b
image	image-klein-base
image	image-klein-base-9b
image	image-klein-shared
image	image-flux2-dev
image	image-flux1-dev
image	image-bonsai-binary
image	image-bonsai-ternary
image	image-zimage-nano
image	image-zimage-max
image	image-zimage-base
image	image-hidream-o1
image	image-hidream-o1-dev
image	image-sensenova-u1-5-8b-mot
image	image-krea2-raw
image	image-krea2-turbo
image	image-qwen-edit-2511
image	image-qwen-edit-2511-lightning
image	image-ideogram4-sdnq-uint4
text-chat	text-chat-mebot
text-chat	text-chat-psi-agent
text-chat	text-chat-gemma4
text-chat	text-chat-diffusiongemma-26b-optiq-4bit
text-chat	text-chat-gemma4-turbo
text-chat	text-chat-gemma4-12b
text-chat	text-chat-gemma4-12b-4bit
vision-chat	vision-chat-gemma4-12b
text-chat	text-chat-gemma4-nano
text-chat	text-chat-gemma4-max
text-chat	text-chat-laguna-s-2-1
text-chat	text-chat-laguna-xs-2-1
text-chat	text-chat-inkling-small
vision-chat	vision-chat-muse-glimmer-30b
text-chat	text-chat-nemotron-35-lightning
omni-chat	omni-chat-nemotron3-nano-30b-a3b-bf16
text-chat	text-chat-q36-nano
vision-chat	vision-chat-q38-27b
vision-chat	vision-chat-q38-27b-4bit
vision-chat	vision-chat-q38-flash-next-mixed
vision-chat	vision-chat-q38-flash-next-3bit
vision-chat	vision-chat-q38-flash-next-3bit-native-ple
vision-chat	vision-chat-q38-flash-next-4bit
text-chat	text-chat-bonsai-27b-1bit
text-chat	text-chat-bonsai-27b-2bit
text-code	text-agent-ornith-9b
vision-chat	text-agent-ornith-35b-mlx-4bit
text-code	text-agent-ornith-35b-mlx-6bit
text-code	text-agent-ornith-35b-mlx-8bit
text-code	text-agent-ornith-35b-mlx
vision-chat	vision-chat-ornith-35b
text-code	text-agent-qwen35-9b
text-code	text-code-north-mini
text-code	text-agent-ornith-35b
text-chat	text-chat-q36-nano-gguf
text-chat	text-agent-deepseek-v4-flash
text-chat	text-chat-lfm25-a1b-8bit
text-chat	text-chat-lfm25-a1b-bf16
text-chat	text-chat-lfm25-1.2b-bf16
text-chat	text-chat-lfm25-1.2b-qad-4bit
text-chat	text-chat-lfm25-2.6b-4bit
text-chat	text-chat-lfm25-2.6b-qad-4bit
text-chat	text-chat-lfm25-2.6b-bf16
vision-chat	vision-chat-lfm25-3b-8bit
speech-tts	speech-tts-qwen3-nano
speech-tts	speech-tts-qwen3-customvoice
speech-asr	speech-asr-qwen3
speech-asr	speech-asr-parakeet
speech-diarization	speech-diarization-sortformer
text-code	text-code-qwen3
text-embed	text-embed-qwen3-0.6b
vision-embed	vision-embed-qwen3-vl-2b
text-anonymize	text-anonymize-privacy-filter
vision-ocr	vision-ocr-infinity-pro
vision-ocr	vision-ocr-infinity-pro-int8
vision-ocr	vision-ocr-lighton
vision-segment	vision-segment-sam31
vision-ground	vision-ground-falcon-perception
vision-flood	vision-flood-terramind-base
vision-face	vision-face-buffalo-l
vision-geometry	vision-geometry-moge2-small
vision-depth	vision-depth-vda-small
vision-depth	vision-depth-vda-small-metric
vision-geometry	vision-geometry-da3-small
image-3d	image-3d-triposr
image-3d	image-3d-instantmesh-base
image-3d	image-3d-trellis2-4b
music	music-acestep
music	music-acestep-xl-base
music	music-acestep-xl-sft
music	music-acestep-xl-turbo
music	music-acestep-xl-turbo-lm4b
music	music-acestep-lm-1.7b
music	music-acestep-lm-4b
music	music-minimax-music3
music	music-magenta-rt2-small
music	music-magenta-rt2-base
music	music-muscriptor-small
music	music-muscriptor-medium
music	music-muscriptor-large
music	music-separate-bs-roformer-viperx-1297
music	music-separate-bs-roformer-4stem
music	music-separate-mel-roformer-dereverb
music	music-separate-mel-roformer-denoise
audio	audio-enhance-ap-bwe-16kto48k
audio	audio-enhance-universr-audio
sfx	sfx-woosh-dflow
sfx	sfx-woosh-flow
sfx	sfx-woosh-clap
sfx	sfx-woosh-synchformer
sfx	sfx-woosh-vflow-8s
sfx	sfx-woosh-dvflow-8s
sfx	sfx-mmaudio-large-44k-v2
video	video-ltx-av
video	video-ltx23-av-mlx
video	video-ltx23-full-mlx
video	video-ltx23-a2vid-mlx
video	video-ltx25-distilled-bf16
video	video-ltx25-full-bf16
video	video-wan22-ti2v-5b-mlx
video	video-minimax-h3-fl2va-mlx
video	video-minimax-h3-fl2va-bf16-mlx
video	video-minimax-h3-fl2va-8bit-mlx
video	video-minimax-h3-fasth3-vsa-datafree-mlx
video	video-minimax-h3-ref2va-mlx
video	video-cosmos3-edge-mlx
video	video-scail2-14b-mlx
video	video-dreamx-world-5b-ar-mlx
vision-fire	vision-fire-terramind-base
vision-embed	vision-embed-tessera-v2-nano
vision-embed	vision-embed-tessera-v2-small
vision-embed	vision-embed-tessera-v2-medium
vision-embed	vision-embed-tessera-v2-large
vision-embed	vision-embed-tessera-v2-teacher
vision-embed	vision-embed-olmoearth-v12-nano
vision-embed	vision-embed-olmoearth-v12-tiny
vision-embed	vision-embed-olmoearth-v12-small
vision-embed	vision-embed-olmoearth-v12-base

Restricted model downloads

mere.run is free and open source, but model and component licenses remain the terms of their respective owners. mere.run does not bundle these weights or decide whether your intended use qualifies. Before downloading an access-gated model or one with a material use limit, such as non-commercial, research-only, or revenue-limited terms, review the listed terms and pass --accept-model-license. Passing the flag and continuing with the download confirms that you accept those terms and agree to comply with them:

mere.run model pull vision-face-buffalo-l --accept-model-license
mere.run model pull --all --accept-model-license

Without that flag, a single-model pull and its preflight are blocked; --all skips restricted models. Restricted models never auto-download from an inference command. The macOS app presents the same explicit acceptance before a download and exposes the term links in its Models pages. Batch downloads through agent onboard --pull-recommended skip restricted entries without acceptance, while open-webui quickstart --pull validates all configured models before downloading any; both accept the same --accept-model-license flag.

Models	Upstream terms that require acceptance
image-flux1-dev	FLUX.1 dev Non-Commercial License v1.1.1; non-commercial, non-production use
image-flux2-dev	FLUX Non-Commercial License and BFL Acceptable Use Policy; non-commercial, non-production use
image-klein-9b, image-klein-base-9b	FLUX Non-Commercial License v2.1; non-commercial, non-production use
image-krea2-raw, image-krea2-turbo	Krea 2 Community License; commercial use is limited to entities below USD 1M trailing annual revenue, plus use/distribution conditions
image-sensenova-u1-5-8b-mot	Apache License 2.0
image-ideogram4-sdnq-uint4	Ideogram Non-Commercial Model Agreement
LFM2.5 text and vision targets plus their DSpark companions	LFM Open License v1.0; commercial use by entities at or above USD 10M annual revenue is excluded
vision-chat-muse-glimmer-30b	Apache-2.0 plus Meta's bundled usage policy; upstream says the model is not intended for download or use by people under 18
vision-chat-q38-flash-next-mixed, vision-chat-q38-flash-next-3bit, vision-chat-q38-flash-next-3bit-native-ple, vision-chat-q38-flash-next-4bit	Qwen Community License 1.0; redistribution, attribution, and restricted-use terms
vision-segment-sam31	Meta SAM License custom use, trade-control, attribution, and redistribution conditions
vision-face-buffalo-l	InsightFace pretrained weights; non-commercial research use
vision-embed-olmoearth-v12-{nano,tiny,small,base}	OlmoEarth Artifact License; prohibited military, defense, intelligence, human-surveillance, policing, and listed extractive uses
video-minimax-h3-fl2va-mlx, video-minimax-h3-fl2va-bf16-mlx, video-minimax-h3-fl2va-8bit-mlx, video-minimax-h3-fasth3-vsa-datafree-mlx, video-minimax-h3-ref2va-mlx	MiniMax-H3 Community License; use, distribution, and display are excluded in the United States, European Union, United Kingdom, and Republic of Korea, with downstream notice and safeguard obligations
image-3d-trellis2-4b	the mounted DINOv3 encoder is gated under Meta's custom DINOv3 License
music-muscriptor-{small,medium,large}	CC BY-NC 4.0 model weights
sfx-woosh-*	CC BY-NC 4.0 Woosh or MMAudio Synchformer weights
sfx-mmaudio-large-44k-v2	CC BY-NC 4.0 MMAudio checkpoints plus Apple's research-only DFN5B encoder terms
video-ltx-av, video-ltx23-av-mlx, video-ltx23-full-mlx, video-ltx23-a2vid-mlx, video-ltx25-distilled-bf16, video-ltx25-full-bf16	LTX-2 Community License; entities at or above USD 10M annual revenue need a paid commercial license, plus acceptable-use conditions. The 2.3 MLX paths also install a hidden Gemma 3 text encoder under Google's Gemma Terms and Prohibited Use Policy. The 2.5 checkpoints include Gemma 4 weights under Apache License 2.0.

The catalog pins every restricted download source to an immutable commit. New managed installs write those repository revisions, every applicable model/component license and URL, and the acceptance result into schema 3 of mererun_model.json. mere.run model info MODEL displays the same record. Pre-existing installs remain usable and are not retroactively treated as an acceptance.

The flag is not a generic click-through for every custom model license. Public, ungated downloads whose licenses grant commercial use by exercising the licensed rights do not require it. This includes Poolside Laguna S/XS 2.1 and NVIDIA Cosmos3-Edge under OpenMDW-1.1, NVIDIA Sortformer under the NVIDIA Open Model License, and the hidden Gemma 3 companion under the Gemma Terms of Use. Their terms still apply. Managed downloads retain the available license, README, attribution, and immutable source provenance.

music-separate-bs-roformer-viperx-1297, music-separate-bs-roformer-4stem, music-separate-mel-roformer-dereverb, and music-separate-mel-roformer-denoise also do not require separate acceptance. The pinned AEmotion Studio model release includes an explicit MIT LICENSE and an MIT model-card declaration. The managed install retains both files and admits the weights, source configuration, model card, and license only when all four match their model-specific frozen byte counts and SHA-256 digests. The MelBand profiles additionally retain separate dereverb and denoise source configs and exact 913 MB checkpoint hashes.

audio-enhance-ap-bwe-16kto48k does not require separate acceptance. AP-BWE's pinned source repository states that both code and pretrained weights are MIT. The managed public transport snapshot retains the code and weights license files, source config, and the exact official 16→48 kHz checkpoint archive; mere.run verifies all four byte counts and SHA-256 digests before loading.

audio-enhance-universr-audio does not require interactive acceptance, but its two upstream licenses must not be conflated. The native port follows the MIT-licensed woongzip1/UniverSR source at its pinned commit. The separately downloaded official woongzip1/universr-audio checkpoint is CC BY 4.0. The managed install verifies the checkpoint, source configuration, and model card by frozen revision, byte count, and SHA-256 before loading.

image-zimage-nano also does not require separate acceptance. Its canonical Tongyi-MAI/Z-Image-Turbo base is Apache-2.0; the pinned mflux conversion's model-card license label is inconsistent with that canonical source and is not treated as a new restriction on the converted weights.

speech-diarization-sortformer installs the fp16 conversion from mlx-community/diar_streaming_sortformer_4spk-v2.1-fp16 at immutable revision e23e6404bd9859e93edbf94a740eb1c7fc58f12e. Its source checkpoint is NVIDIA's diar_streaming_sortformer_4spk-v2.1, referenced at immutable revision fafaab5faa1617a0ca52d38dd3dc4bd636800d3d; the weights are installed separately and are not vendored in this repository.

Install most catalog IDs from managed Hugging Face sources with mere.run model pull. A small number of legacy and local catalog IDs remain so existing installs and explicit local paths keep working:

• image-klein-shared
• text-chat-mebot
• text-chat-psi-agent

image-klein-shared is an internal shared-component install shape, and the text-chat IDs listed here remain local-path-only until they have public Hugging Face sources.

text-agent-qwen35-9b is a low-memory GGUF coding model. It uses the public Hugging Face source unsloth/Qwen3.5-9B-GGUF and selects Qwen3.5-9B-Q4_K_M.gguf. Its text-code API lane does not expose tool calls; the small Pi setup-agent tier instead uses text-agent-ornith-9b.

text-code-north-mini installs the Unsloth GGUF quant of Cohere Labs' North Mini Code 1.0 at the pinned catalog revision. North Mini Code is a 30B total / 3B active coding MoE with a 256K advertised context window; mere.run uses the North-Mini-Code-1.0-UD-Q4_K_M.gguf file so the model runs through the same native Swift/llama.cpp text code path as the existing Qwen coder. It requires a llama.cpp runtime with cohere2moe architecture support.

text-agent-ornith-9b installs the public sahilchachra/ornith-1.0-9b-optiq-5bpw-mlx snapshot at the pinned catalog revision. Ornith is a DeepReinforce agentic coding model with Qwen3.5 text architecture metadata; mere.run treats this MLX OptiQ quant as a native Qwen-family runtime target for chat, api serve, and setup-agent experiments.

The Ornith 1.5 native family pins the official ornith-ai/Ornith-1.5-35B-A3B-MLX-4bit, -6bit, -8bit, and BF16 ornith-ai/Ornith-1.5-35B-A3B-MLX repositories as text-agent-ornith-35b-mlx-4bit, -6bit, -8bit, and text-agent-ornith-35b-mlx. The 35B-parameter MoE activates about 3B parameters per token, advertises a 262,144-token context, and runs through the native Qwen-family runtime. Runtime auto-download remains disabled. Q6, Q8, and BF16 pulls also install the final safetensors shard and index containing mtp.* from the pinned authoritative base checkpoint as the shared text-agent-ornith-35b-mtp companion. The Q4 lane instead pulls the pinned Sawfwair/Ornith-1.5-35B-A3B-MLX-4bit-Vision-MTP packaging snapshot, which contains its target, vision, and MTP files together. The authoritative ornith-ai/Ornith-1.5-35B-A3B base checkpoint declares the model under the MIT license in its model-card metadata; the MLX conversion repository does not duplicate the license file.

text-agent-ornith-35b-mlx-4bit is the recommended local multimodal lane. Its single packaging snapshot combines the official Ornith-1.5-35B-A3B-MLX-4bit text target with the authoritative base checkpoint's model-00001-of-00016.safetensors vision shard, vision configuration, processor metadata, and Shisa's Ornith-distilled BF16 MTP head. The head's 19 fused tensors are packaged as 785 indexed tensors with every BF16 value preserved. Its Apache-2.0 license, notice, and per-tensor hashes are included under mtp/; the target and vision components retain their upstream MIT declaration. The complete snapshot is about 24.2 GiB and records packaged files in SHA256SUMS. Image requests use target-only decoding because MTP speculation is disabled when image embeddings are present; text and code requests retain verified MTP.

The promotion qualification records the head comparison and compatibility checks. After updating mere.run, run mere.run model pull text-agent-ornith-35b-mlx-4bit --force to replace an older Q4 bundle. Existing complete installs are otherwise preserved.

vision-chat-ornith-35b installs the authoritative full BF16 base checkpoint at immutable revision 10fbf86fed7ecee4a061f8b499a618f46001cac1. Unlike the text-only MLX conversions, the 67.0 GiB snapshot includes Ornith's Qwen3.5 vision tower, processor metadata, and embedded mtp.* tensors. It accepts local images through text chat and the OpenAI-compatible API, retains the 262,144-token advertised context, and is available to the VLM, chat, and code benchmark commands. Runtime auto-download remains disabled; the catalog requires 96 GB unified memory and recommends 128 GB. Use it as the optional quality reference instead of the recommended Q4 lane. The source processor advertises images up to 16,777,216 pixels, but this full-BF16 local lane caps the encoded input at 65,536 pixels to remain below the macOS Metal watchdog.

text-agent-ornith-35b installs DeepReinforce's public deepreinforce-ai/Ornith-1.0-35B-GGUF Q4_K_M file at the pinned catalog revision. It runs through the native Swift/llama.cpp text code path for larger Ornith coding-agent comparisons and uses a 32K runtime context by default to keep local evaluations predictable.

text-agent-deepseek-v4-flash is the preferred managed setup-agent tier on 96 GB+ Apple Silicon Macs, with 128 GB recommended. It pulls the official pure-Q2 0731 imatrix GGUF from antirez/deepseek-v4-gguf at an immutable revision (80.76 GiB, SHA-256 ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0). Mere keeps one full-resident DS4 server, caps the operational context at 32K, uses a 1,024-token prefill chunk, and limits disk KV checkpoints to 8 GiB. Smaller Qwen setup agents are lower-memory alternatives, not upgrades from DeepSeek V4 Flash. Avoid repeatedly unloading and reloading this tier under memory pressure.

text-chat-q36-nano uses the public mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit snapshot. That Hugging Face repo includes an MTP head (mtp.safetensors) for OptiQ serving; mere.run loads that draft head when present, but only uses it for adaptive speculative decode when the effective prompt and context window are long enough. Short-context requests decode with the main chat weights.

vision-chat-q38-27b installs Qwen's official Qwen/Qwen3.8-27B BF16 checkpoint at immutable revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. It is a dense 27B native vision-language model with hybrid linear/full attention, a 262,144-token native context window, image and video weights, and Apache-2.0 terms. mere.run supports text and local-image prompts through the native Qwen-family runtime; video input is not yet exposed. The selected snapshot is 55.59 GB, requires an explicit model pull, and is cataloged for 64 GB unified memory minimum with 96 GB recommended. The pull retains the published generation and processor metadata as well as the license.

vision-chat-q38-27b-4bit installs EigenLabs/Qwen3.8-27B-4bit at immutable revision eda45ab47f465d08d6558f0353a2346e2eb9d5b3. This is the independently reconverted 4-bit/group-64 MLX affine target pinned by the Qwen MLX Fast track. It retains the same Qwen3.8 text, code, image, context, and sampling contracts. The managed pull mounts the track's matching proposal-only 4-bit/group-64 head, morgan/qwen38-27b-mtp-r20k-lr3-q4-g64-q2-rerank at revision fd4a99c590dd6e468c0e2a28168c235e32151a4b, under mtp/. The companion uses a proposal-only 2-bit compact readout with exact 4-bit shortlist reranking. Because the optimized target doesn't contain a vision tower, the managed pull also mounts the official Qwen3.8 configuration, processor metadata, weight index, and the first indexed BF16 shard under vision/ from the pinned official revision above. That shard contains all 333 vision tensors; language tensors in it are not loaded. Qwen's official Apache-2.0 license is retained separately. The sources total about 19.49 GB. Greedy MTP decode is enabled by default for this quantized target because its small-batch Q4 verification is serial-exact; set MERERUN_Q35_MTP_SPECULATION=0 to use target-only decode.

vision-chat-q38-flash-next-mixed installs Sawfwair's immutable Qwen3.8-Flash-Next-MLX-Mixed-2bit revision 0bdf3edf02df271e9898f17a7882e5e6a8feb58a. It applies affine Q2/group-128 to the 48 routed-expert banks, Q4 to eligible core matrices and the n-gram table, and retains embeddings, QSA indexers, routers, vision, and selected MTP paths in BF16. Its 73.10 GB of weights are the managed 128 GB Mac profile, with 112 GB minimum and 128 GB recommended.

vision-chat-q38-flash-next-3bit installs Sawfwair's immutable Qwen3.8-Flash-Next-MLX-Activation-3bit revision c699bd611366cbc441377275bd1b7a6d2e18e1df. The 48 routed-expert banks use fresh BF16-to-Q3/group-64 codes with image and text activation-weighted scale and bias refits. Eligible core, MTP, and vision matrices remain Q4. The n-gram table remains Q4/group-32. The 89.67 GB managed pull is cataloged for 96 GB minimum and 128 GB recommended.

The native qualification on a 128 GiB Apple Silicon host passed 58 of 61 cases by exact expected output. The three misses matched the pinned Q4 output. The sealed 16-case holdout produced identical Q3 and Q4 outputs, including exact output on all eight image and OCR cases. The holdout peak was 62.1 GB for Q3 and 77.2 GB for Q4, with no swap growth. These bounded checks establish Q4 no-regression for the tested cases, not general model quality or BF16 parity.

vision-chat-q38-flash-next-3bit-native-ple installs the complete Sawfwair/Qwen3.8-Flash-Next-MLX-Activation-3bit-Native-PLE revision 1cee9301c745836e0abb8933e89cf27a38b98125. It preserves the Q3 model payload and adds a checksum-bound placement manifest. No user conversion or model optimize command is required. If the configured model store is on an external volume, mere.run copies and verifies the 33 PLE-bearing safetensors (32.43 GB) into its internal model cache on first load, writes a PLE-only safetensors index, and reuses that cache on later loads. The remaining model files stay in the configured model store. If the model is already on the internal volume, the runtime reads the PLE table there without making a copy. The cache requires enough free internal space to retain an 8 GiB reserve; otherwise the runtime uses the original model-store files.

vision-chat-q38-flash-next-4bit installs the companion immutable Sawfwair/Qwen3.8-Flash-Next-MLX-4bit revision 6cc9bbc0fae9ce26b7670b3ed1e26d557c154506. Its 104.74 GB of weights use affine Q4/group-64 for eligible matrices, with Q4/group-32 for the 160-wide n-gram table. It is cataloged for 160 GB minimum and 192 GB recommended rather than as a 128 GB alternative.

All three artifacts include the Qwen Community License 1.0 and a complete source/output hash receipt. Pulls are explicit and require either --accept-model-license or its equivalent --accept-license-terms alias. The native Qwen4Exp runtime supports text generation; image input remains unqualified. Version 0.46.0 and later implement the trained QSA micro-block selector beyond 2,048 tokens, replacing v0.45.0's prompt-plus-generation cap. Up to the indexer budget, standard causal attention is exact because every visible block is selected. Beyond it, each query selects 512 complete four-token blocks and includes the current partial block. Future blocks are excluded before selection. The same QSA path is used by the bundled one-layer MTP head, with exact target verification of proposed tokens.

--context-size bounds prompt plus generation up to the checkpoint's published 262,144-token limit. This is an architectural limit, not a 262k memory or quality qualification on a 128 GB Mac. KV and MTP history still grow with context even though sparse-attention temporaries are tiled; an explicit --context-size 32768 keeps the requested window smaller. Existing model downloads already contain the indexer weights, so no new quantization or download is needed.

The mixed checkpoint measured 1.43–1.78x faster MTP decode in the original short-context real-checkpoint comparisons; these are not long-context timings. Set MERERUN_Q35_MTP_SPECULATION=0 to retain target-only decode for comparison or memory pressure.

text-chat-bonsai-27b-1bit and text-chat-bonsai-27b-2bit install Prism ML's public prism-ml/Bonsai-27B-mlx-1bit and prism-ml/Ternary-Bonsai-27B-mlx-2bit snapshots at exact catalog revisions. They are dense Qwen3.6 27B vision/reasoning models with packed affine 1-bit or 2-bit language weights and dense vision weights. The snapshots are approximately 5.13 GB and 8.52 GB respectively and advertise a 262,144-token context. mere.run uses the native Qwen-family text and vision runtime, native low-bit linear and embedding kernels, and the published thinking and sampling defaults. Both models are Apache-2.0 licensed; managed pulls retain upstream license and notice files.

text-chat-q36-nano-gguf installs the Unsloth Qwen3.6-35B-A3B-UD-Q4_K_M.gguf quant. It is the llama.cpp/GGUF companion to the Apple Silicon MLX text-chat-q36-nano path and is the default chat model for Linux CUDA hosts.

vision-chat-muse-glimmer-30b pins Sawfwair's 21.38 GB selective MLX Q4 artifact at revision 6532e898dc5c1a55b51b1b108cd36728b79be751. Its conversion receipt pins Meta's Apache-2.0 Muse Glimmer 30B BF16 source at revision f84ecc3a0ea984a4c04542a84269e3d065350a6e, records every source and output hash, and retains Meta's LICENSE and USAGE_POLICY.md. The managed artifact is never downloaded implicitly. The native Swift/MLX runtime implements its 52-layer local/local/local/global text stack, NoPE global layers, gated attention, 50-layer perception encoder, interleaved image tokens, ATEM tool calls, and low/medium/high/xhigh reasoning-strength prompt contract. Its image path uses the released uint8 Lanczos resize behavior and float32 position interpolation. The released artifact and offline converter use selective Q4/group-64 over 420 text/output/adapter matrices while retaining the token embedding and complete vision tower in BF16; pass --quantization-scope compact to quantize all 721 eligible matrices for an explicit lower-memory experiment. The runtime discovers quantized modules from their .scales arrays, so both layouts use the same inference implementation. The same pull installs z-lab's 5.54 GB DFlash2 assistant as the hidden managed model vision-chat-muse-glimmer-30b-dflash2, pinned to z-lab/Muse-Glimmer-30B-DFlash2@b54ffdd11fa9cfe2af370012e5763d492c904128. The earlier 5.11 GB assistant from meta-models/Muse-Glimmer-30B-assistant at revision 2c86316d689027b91123638739743fef1d425233 remains a compatible fallback for existing installations. Pulls require explicit review and acceptance of the target's bundled LICENSE and USAGE_POLICY.md; upstream says the model is not intended for download or use by people under 18. Python conversion scripts are offline artifact tooling only and are not part of inference.

text-chat-nemotron-35-lightning pins Sawfwair's native MLX conversion at revision 6699e5fd3f0c5b392bb3f8bac2443276bb41958a, produced from NVIDIA's NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 source revision e0b753dc24903ad4d62f5696077da22020eca89a. The receipt records the 52 pinned source shards and every output hash. It repacks the released ModelOpt NVFP4 nibbles bit-for-bit into MLX's native container, retains both block and global scales, and materializes the 46 released FP8 projections as BF16 without a second quantizer. Managed pulls also install the separate Sawfwair MLX conversion of NVIDIA's 967M-parameter DSpark companion at artifact revision d30f0914d6bbb6da36302bd9228f92824901e675, pinned from source revision e3af76fbff445ef795958bee96bc1126af70fd57. Both artifacts retain NVIDIA's OpenMDW-1.1 license and upstream model cards, never download implicitly, and do not require an additional mere.run acceptance gate solely because the public repositories use a custom license identifier.

omni-chat-nemotron3-nano-30b-a3b-bf16 stages one standalone native MLX checkpoint derived losslessly from NVIDIA's Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 source revision 24e67ea000b7c2837fc8f9488aa2008524fac8ba. The native artifact is pinned at revision 9256528c67910cc390c3286157b1527eac44ef04. Its approximately 66 GB download stores every BF16 parameter once: 17 slim shards retain the non-expert tensor keys and payload bytes, while one published shard stacks the routed experts in the physical layout used by MLX. No first-run conversion or second 58.75 GB expert cache is required.

The checkpoint accepts text, images, audio, and video and produces reasoning text. Its document examples render PDF pages to images before prompting; PDF is not a separate raw input modality. The catalog pins the native artifact, byte-level conversion receipt, configuration contracts, tokenizer and media processor metadata, NVIDIA Open Model Agreement terms, 131,072-token composed context, and 20,480-token maximum output.

The native Swift/MLX runtime implements the C-RADIO vision tower, Parakeet audio tower, BF16 Nemotron-H language backbone, multimodal prefill, local video sampling, and media-aware OpenAI chat completions. It is supported on Apple Silicon machines with at least 112 GB unified memory and recommended at 128 GB. The model remains an explicit pull because its NVIDIA terms must be accepted:

mere.run model pull omni-chat-nemotron3-nano-30b-a3b-bf16 \
  --accept-model-license
mere.run text chat \
  --model omni-chat-nemotron3-nano-30b-a3b-bf16 \
  --video clip.mp4 \
  --prompt "Summarize the clip."

The runtime consumes the published native shards in place. A lightweight managed-model wrapper may point at an immutable snapshot on external storage, keeping it visible to model list, text chat, and api serve without duplicating model weights.

text-chat-gemma4-12b and vision-chat-gemma4-12b share Google's dense Gemma 4 12B-it checkpoint; the text id uses the native chat path, while the vision id enables OpenAI image content parts through api serve. Pulling either managed 12B id also pulls the companion google/gemma-4-12B-it-assistant MTP drafter. The native Swift runtime uses that assistant only for greedy decode-tail speculation after text, image, or audio prefill has produced target hidden state and shared KV; raw local model paths and sampled generations fall back to baseline decode. text-chat-gemma4 is the dense bf16 Gemma 4 31B alias and is gated for larger machines. On 32 GB Apple Silicon Macs, use text-chat-gemma4-turbo, which installs the MLX NVFP4 Gemma 4 26B-A4B-it MoE snapshot and runs through the native Swift Gemma runtime. On smaller supported machines, text-chat-gemma4-12b-4bit installs the checksum-pinned Sawfwair/gemma-4-12B-it-MLX-4bit snapshot as the compact managed Gemma 12B chat tier. Sawfwair's package is independently converted from Google's verified dense checkpoint, includes MERERUN_CONVERSION.json with source and emitted artifact hashes, and is published as the user-facing v1.0.0 release.

text-chat-diffusiongemma-26b-optiq-4bit pins the public Apache-2.0 mlx-community/diffusiongemma-26B-A4B-it-OptiQ-4bit snapshot at revision 30f3c7c7746bf41cfd1a290155cc3b777ab588b9. The 17.85 GB snapshot stores the ordinary projections and shared embedding in mixed 4-bit and 8-bit affine MLX tensors. Its separate optiq/optiq_vision.safetensors file retains the vision tower in BF16. The native Swift runtime consumes the OptiQ tensors in place, uses a causal prompt cache plus a bidirectional denoising canvas, and supports text and function-tool output with at most 256 generated tokens. The text runtime validates the vision sidecar but rejects image input because that path has no real-checkpoint qualification.

text-chat-lfm25-a1b-8bit uses the public LiquidAI/LFM2.5-8B-A1B-MLX-8bit snapshot at the pinned catalog revision. It is a text-only MLX 8-bit directory-root model with config.json, tokenizer.json, tokenizer_config.json, and sharded *.safetensors weights. mere.run runs it through the native Swift LFM2 runtime; no Python bridge is used. This quantized target does not attach the BF16-trained DSpark assistant.

text-chat-lfm25-a1b-bf16 uses LiquidAI/LFM2.5-8B-A1B at revision b9aebfcbe28b6cb374042f495d733037550ab146, the exact target for the managed 8B-A1B DSpark companion.

text-chat-lfm25-1.2b-bf16 uses LiquidAI/LFM2.5-1.2B-Instruct at revision df58c174f05ff733f83f8cae10ea9298224c8006. Its matching text-chat-lfm25-1.2b-dspark companion uses LiquidAI/LFM2.5-1.2B-Instruct-DSpark at revision 4876d04848e15a6fd48d7c1481110e7cf5d62621.

text-chat-lfm25-1.2b-qad-4bit uses Sawfwair/LFM2.5-1.2B-Instruct-QAD-MLX-4bit at revision b1caaa502c5bfd975f5219162eaa90b1e7cc7839. The conversion pins LiquidAI's QAD Q4_0 GGUF at revision afbd8eaeab5dd94ba0b079ebfb02517d19641e38, keeps every projection nibble and scale exact in MLX group-32 affine tensors, and records all source and output hashes in MERERUN_CONVERSION.json.

text-chat-lfm25-2.6b-4bit uses the pinned 4bit/ partition of LiquidAI/LFM2.5-2.6B-MLX. The managed pull selects only that partition plus the repository license and model card, then normalizes the nested directory for the native dense Lfm2ForCausalLM runtime. The checkpoint uses affine 4-bit linear weights with a 6-bit tied embedding and is approximately 1.60 GB. It remains target-only because the managed DSpark assistant was trained for the unquantized target.

text-chat-lfm25-2.6b-bf16 uses LiquidAI/LFM2.5-2.6B at revision a334ee78cd38458bb71eda24109ac42dcec1309d, the exact target for the managed 2.6B DSpark companion.

text-chat-lfm25-2.6b-qad-4bit uses Sawfwair/LFM2.5-2.6B-QAD-MLX-4bit at revision e829e540629759fdb886cb529e4d03640a11ffa7. Its conversion pins LiquidAI's QAD Q4_0 GGUF at revision f4a289c8a200a5ca71005ba7abc2dad33058a450 and follows the same exact projection conversion contract. The standard 2.6B MLX 4-bit checkpoint remains the faster native MLX lane; the QAD variant exists for the quantized-accuracy tradeoff. Liquid's published QAD throughput values reuse native llama.cpp Q4_0 throughput because QAD and PTQ share that format, not because QAD itself adds a runtime speedup.

The three exact BF16 LFM2.5 text targets install matching DSpark sidecars from LiquidAI: LFM2.5-8B-A1B-DSpark at revision 5b285c827912834665b1915f171897e49ff0f388, LFM2.5-2.6B-DSpark at revision 458cedab07d0f7b2b05700c77e1aa463d43d6f04, and the 1.2B revision above. Each sidecar is roughly 0.59-0.66 GB and is loaded only beside its compatible target. LiquidAI reports mean M4 Max speedups of 2.54x for 1.2B, 2.27x for 2.6B, and 1.18x for 8B-A1B; these are upstream measurements, not mere.run benchmark results. See LiquidAI's LFM2.5-DSpark release and methodology for the upstream benchmark protocol and framework integrations.

vision-chat-lfm25-3b-8bit uses the public LiquidAI/LFM2.5-VL-3B-MLX-8bit checkpoint at revision 4065d2c056a9c54d44fec67cf651812b55c6673f. The managed snapshot is approximately 3.74 GB and includes the dense LFM2.5 2.6B language backbone, SigLIP2 NaFlex vision tower, multimodal projector, tokenizer, chat template, and processor_config.json. mere.run processes local file paths and base64 data URLs natively, expands each <image> placeholder to the downsampled patch grid, and continues generation through the shared LFM2 decode engine. Remote image URLs are not fetched by the local runtime.

Useful environment variables for that path:

• MERERUN_HUB_CACHE: override the native Hugging Face snapshot cache path

image-klein-9b installs the ungated mlx-community/FLUX.2-klein-9B mirror for larger distilled Klein generation and reference-image workflows. It is distinct from image-klein-base-9b, which pulls the undistilled Base 9B transformer plus shared 9B components for higher-capacity LoRA training and research workflows.

image-flux1-dev installs the gated Diffusers layout from black-forest-labs/FLUX.1-dev at revision 3de623fc3c33e44ffbe2bad470d0f45bccf2eb21. The approximately 34 GB managed download includes the FLUX.1 transformer, CLIP-L and T5-XXL text encoders, tokenizers, VAE, and flow-matching scheduler. The managed patterns exclude the duplicate root checkpoint files.

Before you pull the model, accept its access terms on Hugging Face and configure a Hugging Face token in mere.run. The pull also requires explicit local acceptance of the FLUX.1 dev Non-Commercial License v1.1.1:

swift run mere.run model pull image-flux1-dev --accept-model-license
swift run mere.run image generate \
  --model image-flux1-dev \
  --prompt "TRIGGER_TOKEN a portrait in a glass conservatory" \
  --lora ./flux1-subject.safetensors=0.8 \
  --output ./flux1-subject.png

The runtime defaults to 28 denoising steps and embedded guidance scale 3.5. Repeat --lora PATH_OR_ID[=SCALE] to apply ordered FLUX.1 adapters with independent scales. Each adapter must target FLUX.1-dev. FLUX.2 and Klein adapters use different transformer dimensions and fail tensor-shape validation.

The model license limits the weights and derivatives to non-commercial, non-production use. It also requires filters or manual review for generated content. The mere.run acceptance record doesn't determine whether a workflow complies with those requirements.

image-flux2-dev installs the gated FLUX.2-dev transformer, variational autoencoder (VAE), tokenizer, and scheduler from black-forest-labs/FLUX.2-dev. The install mounts a pinned 4-bit MLX conversion of the matching Mistral Small 3.2 24B text encoder. This layout reduces the managed download from approximately 113 GB to 78 GB without changing the transformer that a FLUX.2-dev LoRA targets.

Before you pull the model, accept its access terms on Hugging Face, and configure a Hugging Face token in mere.run. The pull also requires explicit local acceptance because the FLUX Non-Commercial License limits the weights to non-commercial, non-production use, and the BFL Acceptable Use Policy applies:

swift run mere.run model pull image-flux2-dev --accept-model-license
swift run mere.run adapter pull flux2-dev-turbo-8step --accept-license
swift run mere.run image generate \
  --model image-flux2-dev \
  --prompt "a ceramic vessel on dark linen, precise studio lighting" \
  --lora flux2-dev-turbo-8step=1.0 \
  --lora ./flux2-dev-style.safetensors=0.8 \
  --output ./flux2-dev-style.png

The managed Turbo adapter is the fal/FLUX.2-dev-Turbo artifact pinned at revision 9ee51cd87578162cf8d02355a870bc5f4570045c. mere.run verifies its 2,760,818,216-byte safetensors file against SHA-256 f76cf9c2cc546ddca878799136434a1098477af3f4b0adff2cfd79f2ebe4aa01. The separate adapter pull requires explicit acceptance because the adapter inherits the FLUX.2-dev noncommercial terms and BFL Acceptable Use Policy.

Selecting flux2-dev-turbo-8step applies the publisher's eight-step, guidance-2.5 sigma recipe. Repeated --lora PATH_OR_ID[=SCALE] values preserve their order and combine the Turbo update with local FLUX.2-dev style or subject adapters. FLUX.1-dev adapters are structurally incompatible and are rejected by the runtime's tensor-shape validation.

The license treats generated outputs separately and permits personal, scientific, and commercial output use subject to its terms. It also requires filters or manual review for generated content. The mere.run acceptance record does not provide legal clearance or certify that a workflow satisfies those requirements.

The default profile uses 50 denoising steps and embedded guidance scale 4.0. The capability catalog requires at least 96 GB of unified memory and recommends 128 GB. Real-checkpoint qualification remains separate from catalog and synthetic runtime tests.

image-bonsai-binary and image-bonsai-ternary map to PrismML Apple Silicon Bonsai Image snapshots:

• prism-ml/bonsai-image-binary-4B-mlx-1bit
• prism-ml/bonsai-image-ternary-4B-mlx-2bit

The snapshot uses a FLUX.2 Klein transformer, but its component names are upstream-specific. Managed or local roots are expected to contain:

• manifest.json
• tokenizer/tokenizer_config.json
• text_encoder-mlx-4bit/config.json
• text_encoder-mlx-4bit/model.safetensors or model.safetensors.index.json
• transformer-packed-mflux/config.json
• transformer-packed-mflux/quantization_config.json
• transformer-packed-mflux/diffusion_pytorch_model.safetensors
• vae/config.json
• vae/diffusion_pytorch_model.safetensors
• scheduler/scheduler_config.json

The binary manifest records the transformer as 1-bit g128 Prism packed affine weights; the ternary manifest records 2-bit g128 MLX packed affine ternary weights. Both keep the text encoder in the upstream 4-bit MLX layout and run generation through the native Swift FLUX.2 Klein pipeline with four steps, CFG 1.0, and sigma shift 3.0 by default. The binary runtime path uses native packed 1-bit affine matmul kernels on Metal and Linux CUDA, with a dequantized MLX fallback for non-GPU or unsupported shapes.

vision-segment-sam31 packages the native SAM 3.1 segmentation and tracking runtime used by mere.run vision segment, mere.run vision track, and mere.run vision track-live. Managed or local SAM roots are expected to contain:

• config.json
• model.safetensors or model.safetensors.index.json
• tokenizer/tokenizer.json
• tokenizer/tokenizer_config.json

Tokenizer files are technically optional for geometry prompts, but managed installs include them because text prompts are the preferred segmentation path. The managed package mounts tokenizer assets from the SAM 3.1 mirror while the native MLX weights come from mlx-community/sam3.1-bf16.

The manifest for this package advertises both vision_segmentation and vision_tracking capabilities.

image-hidream-o1 and image-hidream-o1-dev map to the public HiDream O1 image checkpoints:

• HiDream-ai/HiDream-O1-Image
• HiDream-ai/HiDream-O1-Image-Dev

HiDream O1 uses a unified pixel-transformer root layout rather than a VAE/text-encoder component tree. Managed or local roots are expected to contain:

• config.json
• tokenizer_config.json
• tokenizer.json or vocab.json plus merges.txt
• preprocessor_config.json
• model.safetensors or model.safetensors.index.json

The native Swift runtime validates this layout, decodes the typed root configuration, prepares text/reference sample metadata, and runs generation through the downloaded Qwen3-VL decoder, vision tower, timestep embedder, patch embedder, generation-aware attention mask, and HiDream pixel head. Text-only generation, one-reference instruction editing, and multi-reference subject personalization share the same native path; reference modes additionally run Qwen3-VL vision preprocessing and replace chat-template image placeholders before denoising.

Runtime defaults come from the managed manifest:

• image-hidream-o1-dev: 28 steps, CFG 0.0, fixed flash FlowMatch schedule
• image-hidream-o1: 50 steps, CFG 5.0, shifted Flow UniPC schedule

Both checkpoints are large BF16 unified-transformer roots, about 33 GiB on disk each before filesystem compression effects. Expect high unified-memory pressure and prefer one-step smokes before full-quality 28/50 step runs.

image-sensenova-u1-5-8b-mot maps to the official sensenova/SenseNova-U1.5-8B-MoT checkpoint at the exact revision recorded in the managed catalog. The corresponding reference implementation is published at OpenSenseNova/SenseNova-U1. The root contains the typed configuration, Qwen tokenizer assets, a safetensors index, and 13 shards.

SenseNova uses one unified checkpoint rather than separate text-encoder and VAE directories. The native runtime mirrors the checkpoint's understanding and generation experts, caches the prompt/reference prefix, embeds RGB patches, applies the published shifted Euler flow schedule, and decodes predicted RGB pixels with the checkpoint's convolutional pixel head. Text-to-image and multi-reference instruction editing share this Swift/MLX path. The managed defaults are 50 steps, CFG 4.0, and timestep shift 3.0.

image-krea2-raw and image-krea2-turbo map to Krea's public Krea 2 checkpoints:

• krea/Krea-2-Raw
• krea/Krea-2-Turbo

Managed or local roots are expected to contain the component Diffusers layout:

• model_index.json
• tokenizer/tokenizer_config.json
• tokenizer/tokenizer.json
• text_encoder/config.json
• text_encoder/model.safetensors
• transformer/config.json
• transformer/diffusion_pytorch_model.safetensors.index.json
• transformer/diffusion_pytorch_model-*.safetensors
• vae/config.json
• vae/diffusion_pytorch_model.safetensors
• scheduler/scheduler_config.json

Krea also publishes root-level raw.safetensors and turbo.safetensors transformer copies for the official codebase. Managed pulls intentionally exclude those files and pull the split transformer component instead, so the model store does not download the same large transformer payload twice.

The native Swift runtime follows the public Krea 2 sampler shape: Qwen3-VL text conditioning with the Krea system prefix, layer-selected hidden-state fusion, single-stream MMDiT denoising, 16-pixel image-token alignment, FlowMatch Euler steps, and Qwen Image VAE decoding. The wired public generation mode is text-to-image with optional LoRA adapters; reference images and image-to-image are not supported for this family yet. LoRA training uses image-krea2-raw; inference uses image-krea2-turbo.

Krea publishes sample LoRA adapters such as krea/Krea-2-LoRA-retroanime and krea/Krea-2-LoRA-kidsdrawing. Those adapters are trained on Raw and loaded on Turbo with Diffusers lora_A / lora_B keys. The native loader preserves Krea's img_in, txt_in, text_fusion, time_embed, time_mod_proj, transformer_blocks, and final_layer module names so those published adapters can be used as compatibility references.

Runtime defaults come from the managed manifest:

• image-krea2-raw: 52 steps, CFG 3.5, FlowMatch shift/mu 1.15
• image-krea2-turbo: 8 steps, CFG 0.0, FlowMatch shift/mu 1.15

The component install is about 36 GiB before filesystem compression effects. Expect high unified-memory pressure and prefer 96 GB+ Apple Silicon machines.

image-qwen-edit-2511 is the immutable quality lane for Qwen Image Edit 2511. It pulls Qwen/Qwen-Image-Edit-2511 at revision 6f3ccc0b56e431dc6a0c2b2039706d7d26f22cb9; it does not replace or redirect the legacy qwen-image-edit install. Its managed defaults are 40 steps and true CFG 4.0 with a blank negative prompt.

image-qwen-edit-2511-lightning uses the same pinned base plus lightx2v/Qwen-Image-Edit-2511-Lightning at revision d74eba145674fd7e31b949324e148e21e7118abd. The runtime accepts only the four-step BF16 adapter named Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors, verifies its 849,608,296-byte size and SHA-256 22226e8d05d354bb356627d428809f5afd7819399b077238a2b70a82883a904f, and requires exactly four steps with CFG disabled.

Both managed IDs use the native 2511 edit path. It accepts one to three ordered images across --input followed by repeated --ref-image arguments. Each reference retains its own aspect ratio for Qwen2.5-VL semantic conditioning and VAE appearance conditioning. The output starts from pure noise; packed reference latents remain conditioning tokens rather than being blended into the output latent. These two pinned lanes retain the base encoder and transformer in BF16 at runtime; the legacy qwen-image-edit compatibility path keeps its existing runtime Q4 behavior.

mere.run model pull image-qwen-edit-2511
mere.run image generate \
  --model image-qwen-edit-2511 \
  --input ./primary.png \
  --ref-image ./style.png \
  --prompt "Apply the palette of Picture 2 while preserving Picture 1's layout" \
  --output ./edited.png

The pinned base is about 53.75 GiB before filesystem compression; the Lightning lane adds about 810 MiB. These sizes and the current runtime's memory behavior require physical Apple Silicon qualification before publishing a supported memory class or performance claim.

image-ideogram4-sdnq-uint4 maps to WaveCut's public SDNQ uint4 Ideogram 4 snapshot:

• WaveCut/ideogram-4-sdnq-uint4

Managed or local roots are expected to contain:

• model_index.json
• quantization_manifest.json
• tokenizer/tokenizer_config.json or tokenizer/tokenizer.json
• text_encoder/config.json
• transformer/config.json
• transformer/diffusion_pytorch_model.safetensors
• unconditional_transformer/config.json
• unconditional_transformer/diffusion_pytorch_model.safetensors
• vae/config.json
• vae/diffusion_pytorch_model.safetensors
• scheduler/scheduler_config.json

The managed manifest records SDNQ asymmetric uint4 weights and the separate positive and unconditional transformer branches used by Ideogram 4 guidance. The native runtime can pull, inspect, validate, and decode SDNQ uint4 linear, embedding, and Conv2d weights, build Qwen3-VL concatenated text features, pack Ideogram 4 text/image samples, run positive/unconditional CFG denoising, and decode PNG output through the Flux2-style VAE. Text-to-image image generate is wired; image-to-image, reference inputs, and LoRA are still unsupported for this family.

Locally converted Parakeet Core ML/MLX package

The optional Parakeet Core ML/MLX package is a locally built artifact, not a managed model ID. scripts/model-conversion/convert_parakeet_coreml.py pins the official nvidia/parakeet-tdt-0.6b-v3 source at revision 541d1f99c6b0c3cd0b11a95167540bb8edefd82b and verifies the source checkpoint before conversion. The converted weights remain licensed under CC BY 4.0.

The schema-v3 artifact contains compiled Core ML encoder and decoder models, a decoder embedding table, a compact 13-tensor MLX fallback checkpoint, generated runtime configuration and vocabulary, and the pinned upstream tokenizer. It replaces speech-asr-parakeet for this provider path. Pass its directory explicitly with speech transcribe --backend parakeet --provider coreml --coreml-encoder PATH. Schema-v2 packages use the compact MLX decoder. Legacy schema-v1 encoder-only artifacts use a separate MLX checkpoint.

Hugging Face cache

Hub snapshots use the shared cache location managed by the runtime. Override it when you want large models on an external disk:

export MERERUN_HUB_CACHE=/Volumes/Models/huggingface
swift run mere.run model pull image-zimage-nano

Model pulls are resumable at the file level through the Hugging Face snapshot cache. The CLI writes a managed-model symlink from the mere.run model store to the prepared snapshot when needed.

Hardware support checks

Managed pulls are gated by the local capability catalog before any download. The check uses supported local runtimes plus memory thresholds for each model family, then blocks models that are unlikely to run reliably on this machine.

Inspect the local recommendation first:

swift run mere.run model capabilities
swift run mere.run model capabilities --all

If you are intentionally testing an unsupported setup, pass --allow-unsupported to mere.run model pull.

Model store behavior

The CLI resolves models in this order:

1. --models-root process override
2. MERERUN_MODELS_DIR
3. persisted local model-store setting
4. default ~/Library/Application Support/MereRun/models

Examples:

# Pull into the default model store
swift run mere.run model pull image-zimage-nano

# Pull into a custom SSD-backed store
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-q36-nano
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-lfm25-a1b-8bit --accept-model-license
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-lfm25-a1b-bf16 --accept-model-license
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-lfm25-1.2b-bf16 --accept-model-license
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-lfm25-1.2b-qad-4bit --accept-model-license
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-lfm25-2.6b-4bit --accept-model-license
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-lfm25-2.6b-bf16 --accept-model-license
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull text-chat-lfm25-2.6b-qad-4bit --accept-model-license
MERERUN_MODELS_DIR=/Volumes/Models swift run mere.run model pull vision-chat-lfm25-3b-8bit --accept-model-license

# Inspect installed models
swift run mere.run status
swift run mere.run model list
swift run mere.run model info image-klein-max

Music, SFX, and video layouts

Some retained surfaces have more structure than a flat model root.

ACE-Step 1.5 models

The top-level model roots are:

.../models/music-acestep
.../models/music-acestep-xl-turbo
.../models/music-acestep-xl-turbo-lm4b
.../models/music-acestep-xl-sft
.../models/music-acestep-xl-base
.../models/music-acestep-lm-1.7b
.../models/music-acestep-lm-4b

Those roots may contain:

• acestep-v15-turbo/
• acestep-v15-xl-turbo/
• acestep-v15-xl-sft/
• acestep-v15-xl-base/
• acestep-5Hz-lm-1.7B/ or another supported LM subdirectory
• acestep-5Hz-lm-4B/
• Qwen3-Embedding-0.6B/
• vae/

Older local installs that still use music-acestep-v15-turbo/ remain supported.

music-acestep-xl-turbo pulls the ACE-Step 1.5 XL turbo DiT into acestep-v15-xl-turbo/ and reuses the base ACE-Step 1.5 VAE and Qwen3 text encoder components. music-acestep-xl-turbo-lm4b adds the optional acestep-5Hz-lm-4B/ 5 Hz LM. The independently pullable music-acestep-lm-1.7b and music-acestep-lm-4b planner models can be paired with any ACE-Step DiT through --lm-model; no shared-directory symlink is required. The 1.7B planner is the upstream default. Local component discovery therefore prefers acestep-5Hz-lm-1.7B/ when both sizes are present; select --lm-model music-acestep-lm-4b only when you explicitly want the optional 4B planner. music-acestep-xl-sft and music-acestep-xl-base install the non-distilled XL checkpoints and use continuous flow sampling with CFG, APG, or ADG. Base is also the checkpoint family for extract, lego, and complete tasks. Every component is downloaded at the immutable revision listed in ACE-Step validation.

mere.run music generate, music analyze, and music serve first look for a planner in the selected checkpoint root, then reuse an installed standalone or music-acestep 1.7B planner, and finally pull music-acestep-lm-1.7b when LM planning is required. Override that resolution with --lm-model or the legacy same-root --lm-subdirectory.

music-minimax-music3

MiniMax Music 3 comes from MiniMaxAI/MiniMax-Music3 at immutable revision bd348f9c49ea3c1b39f33ace3436f8fad435f24e. The managed pull selects the 25 runtime files needed by the native Swift/MLX implementation and omits the duplicate qwen_7B/ tree and Python-only examples. The selected snapshot is 28,517,620,807 bytes (approximately 26.6 GiB):

.../models/music-minimax-music3
├── condition_encoder/
├── language_model/
├── rvq_depth_decoder/
├── scheduler/
├── tokenizer/
├── transformer/
├── vocoder/
├── config.json
├── modular_model_index.json
└── LICENSE

The checkpoint uses the MiniMax-Music3 Community License rather than an OSI open-source license. Managed pulls therefore require --accept-model-license, runtime auto-download is disabled, and applications using the model must preserve the upstream attribution and usage restrictions. Review the pinned LICENSE before deploying or redistributing the weights.

mere.run music generate --model music-minimax-music3 assembles the released caption-and-lyrics prompt contract, generates 25 Hz semantic plus residual RVQ codes, runs the overlap-aware flow transformer, and writes native 44.1 kHz stereo WAV output through the released vocoder. --max-frames exposes the upstream 1...9,000-frame contract directly, while --duration converts seconds at 25 Hz. The model card describes supported-quality generation through five minutes; 9,000 frames is the six-minute runtime hard limit.

--minimum-duration and --min-frames add an EOS floor; duration floors round up across the vocoder's whole 512-sample hops so the decoded WAV does not undershoot. --performance-mode optimized is the BF16 default and uses a compact semantic projection, a full-vocabulary sampling view, fused global language-model projections, and batched flow guidance. The eight-token residual depth prefix is recomputed without projection fusion because its cached/fused BF16 path changed seeded codebook choices and caused lyric drift. The opt-in q8-lm and q4-lm apply group-64 affine quantization only to the global language model while retaining the residual-depth decoder in BF16. They are the preferred quantized experiments because depth codebooks directly affect vocal detail. Legacy q8 and q4 still quantize both autoregressive components for compatible maximum compression. Flow stays BF16 because installed-checkpoint timing showed its quantized kernels regress.

--compose runs a local native Gemma4 or Qwen-family chat model in two constrained-JSON passes before loading MiniMax: a bar-aware song blueprint, then lyrics and the checkpoint's three-part structured caption. Supplied lyrics remain authoritative. The typed composition receipt and lyric-duration preflight are stored separately and in schema 7 generation provenance. The recipe and schema 3 profile record resolved language-model and depth-decoder precision independently. Composer weights are unloaded before music inference. The default lyric preflight warns because --duration is an upper bound rather than a promise; strict rejects sparse or structurally invalid inputs without silently adding an EOS floor.

Euler, full autoregressive CFG, and full flow CFG remain the parity defaults. The opt-in --flow-solver ab2, --ar-cfg-frames, and --flow-cfg-end controls support reproducible full-song A/B experiments and are recorded in recipes and profiles. They are experimental controls, not admission claims.

Generation defaults to --memory-mode staged, which releases the language, flow, and vocoder weights between stages. --memory-mode resident keeps the entire stack loaded for repeated work. Staged mode moves the catalog floor to 32 GB unified memory; 64 GB remains recommended for practical song lengths. --sample-rate 32000 produces the same stereo PCM rate exposed by the upstream SGLang speech route; 44,100 Hz remains the native CLI default. music serve --model music-minimax-music3 exposes the non-streaming /v1/audio/speech request shape with input, instructions, seed, and max_new_tokens, plus explicit native duration, step, guidance, solver, CFG-cutoff, lyric-preflight, and sample-rate controls.

MiniMax publishes its optional music-caption-rewriter agent skill separately in the official MiniMax-Music-3 repository. It is not part of the checkpoint snapshot or this distribution. The model-specific mere.run guide music generate --model music-minimax-music3 shows the official install command, the reviewed source commit, and the complete raw Diffusers parameter mapping.

music-magenta-rt2-small and music-magenta-rt2-base

Magenta RT2 models use exported runtime assets from google/magenta-realtime-2 at revision 010aa0dcb0dfd27b24f0ad07b4dad63e8f9521cc. The managed pull keeps only the files needed by the native runtime:

.../models/music-magenta-rt2-small
├── models/mrt2_small/mrt2_small.mlxfn
├── models/mrt2_small/mrt2_small_state.safetensors
├── resources/musiccoca/
└── resources/spectrostream/

The base model uses models/mrt2_base/ with matching mrt2_base filenames. Raw checkpoints/*.safetensors files are not a complete mere.run layout.

mere.run music generate --model music-magenta-rt2-small renders an offline WAV. mere.run music realtime --model music-magenta-rt2-small plays on the default macOS audio device and can capture to WAV with --output.

music-muscriptor-small, music-muscriptor-medium, and music-muscriptor-large

MuScriptor checkpoints come from the gated MuScriptor/muscriptor-{size} Hugging Face repositories. Each managed root contains config.json and model.safetensors. The weights are CC BY-NC 4.0 and require accepting the upstream access terms before mere.run model pull can download them.

mere.run music transcribe decodes input audio to mono 16 kHz, runs the exact published five-second HTK mel frontend and native MLX transformer, then writes a format-1 MIDI file with one track per detected instrument. JSON and JSONL event output are also available.

sfx-woosh-dflow

The Woosh DFlow model root is:

.../models/sfx-woosh-dflow
└── checkpoints/
    ├── Woosh-DFlow/
    ├── Woosh-AE/
    └── TextConditionerA/
        └── tokenizer/

The managed pull uses the AEmotionStudio/woosh-models Hugging Face mirror for Sony Research Woosh v1.0.0 weights and mounts FacebookAI/roberta-large tokenizer files under checkpoints/TextConditionerA/tokenizer/. The native runtime exposes the text-to-SFX distilled DFlow path through mere.run sfx generate.

sfx-woosh-flow

The Woosh original Flow model root is:

.../models/sfx-woosh-flow
└── checkpoints/
    ├── Woosh-Flow/
    ├── Woosh-AE/
    └── TextConditionerA/
        └── tokenizer/

The managed pull uses the same mirror and tokenizer mount as sfx-woosh-dflow. The native runtime exposes the original text-to-SFX Flow checkpoint through mere.run sfx generate --model sfx-woosh-flow; it generally needs more denoise steps than the distilled model.

Woosh CLAP and V2A

sfx-woosh-clap installs checkpoints/Woosh-CLAP/ plus the mounted RoBERTa tokenizer for native text/audio scoring through mere.run sfx clap score.

sfx-woosh-dvflow-8s and sfx-woosh-vflow-8s install the distilled and original video-to-audio checkpoint stacks from AEmotionStudio/woosh-models. Both include checkpoints/Woosh-AE/ and checkpoints/TextConditionerV/.

sfx-woosh-synchformer installs the companion mmaudio_synchformer_fp16.safetensors visual extractor from Kijai/MMAudio_safetensors. mere.run sfx video generate uses it when the input is a raw video file; .npy inputs can still provide precomputed Synchformer synch_out features directly.

sfx-mmaudio-large-44k-v2

The native MMAudio install combines pinned public artifacts into one managed root:

.../models/sfx-mmaudio-large-44k-v2
├── mmaudio_large_44k_v2_fp16.safetensors
├── apple_DFN5B-CLIP-ViT-H-14-384_fp16.safetensors
├── mmaudio_synchformer_fp16.safetensors
├── mmaudio_vae_44k_fp16.safetensors
├── clip/tokenizer.json
└── bigvgan/
    ├── config.json
    └── bigvgan_generator.pt

The MMAudio, CLIP, Synchformer, and VAE safetensors come from the pinned Kijai/MMAudio_safetensors snapshot. CLIP tokenizer/config files are mounted from Apple's pinned DFN5B CLIP repository. BigVGAN-v2 config and generator weights are mounted from NVIDIA's pinned 44.1 kHz repository. The Swift runtime loads the official BigVGAN PyTorch state dictionary with a restricted parser; it does not execute Python or arbitrary pickle globals.

The hkchengrex/MMAudio architecture source is MIT-licensed. The released MMAudio checkpoints are separately CC-BY-NC-4.0 and therefore non-commercial. Apple's mounted DFN5B CLIP model is separately restricted to research purposes under the Apple Machine Learning Research Model License Agreement. NVIDIA's BigVGAN-v2 source and model are MIT-licensed. The managed install retains the exact Apple and NVIDIA license files beside those components. Review all model terms before downloading or using generated output.

video-ltx-av

The unified AV model root is:

.../models/video-ltx-av

mere.run video generate --quality draft --model video-ltx-av is the faster video-only draft path. The legacy --variant selector remains available for older scripts using this root. MERERUN_VIDEO_LTX_MODEL_ROOT can still point at this layout explicitly.

video-ltx23-av-mlx

The LTX 2.3 MLX split model root is:

.../models/video-ltx23-av-mlx

It pulls the distilled split checkpoint from dgrauet/ltx-2.3-mlx, including split_model.json, connector weights, separate video VAE/audio VAE/vocoder files, and the LTX 2.3 upscalers. mere.run model pull video-ltx23-av-mlx also installs the hidden text-encoder-ltx-gemma3-12b-4bit companion used for Gemma 3 prompt conditioning. Set MERERUN_VIDEO_LTX_TEXT_ENCODER_ROOT only when pointing at an external mlx-community/gemma-3-12b-it-4bit checkout.

The native Swift runtime uses this standalone distilled transformer for the fast --quality draft lane. The default --output-mode video-only route retains the checkpoint's joint AV denoising tokens but does not load or decode the audio VAE/vocoder, and its MP4 contains no audio stream. It can also emit synchronized AV with --output-mode audio-video, but the canonical final quality path uses video-ltx23-full-mlx. The Unsloth LTX-2.3-GGUF checkpoint family is a separate quantized GGUF lane and is not loaded by the native MLX video runtime.

video-ltx23-full-mlx

The full LTX 2.3 quality root is:

.../models/video-ltx23-full-mlx

It pulls the full/dev transformer, the official rank-384 distilled LoRA, connector, audio VAE, BWE vocoder, video VAE encoder/decoder, and x2 spatial upscaler from the same immutable dgrauet/ltx-2.3-mlx revision. It does not duplicate the standalone distilled transformer or pull unrelated x1.5 and temporal upscalers. Hugging Face cache objects are shared with video-ltx23-av-mlx when both models are installed. The hidden Gemma 3 companion is shared as well.

The same bundle drives all official two-stage final-quality contracts. The default --quality final --output-mode video-only route runs the full video pipeline without requiring audio in the deliverable. With --output-mode audio-video, unified AV jointly denoises guided video and audio latents in stage one, then refines both after LoRA fusion. A2Vid encodes source audio and freezes those audio latents through both stages; the original decoded source segment—not VAE-decoded audio—is muxed into the MP4.

video-ltx23-a2vid-mlx

This deprecated compatibility ID preserves existing A2Vid installs and scripts:

.../models/video-ltx23-a2vid-mlx

Its legacy narrow manifest may omit the vocoder, so it can run final-quality video-only and source-audio A2Vid but not generated-audio output. Requests for either the legacy ID or the replacement full ID resolve to an already-installed compatible root when possible. Use video-ltx23-full-mlx for pulls.

video-ltx25-distilled-bf16

The official packed LTX 2.5 BF16 root is:

.../models/video-ltx25-distilled-bf16

mere.run model pull video-ltx25-distilled-bf16 --accept-model-license pulls the complete public runtime from Sawfwair/LTX-2.5-Distilled-BF16-MLX-Q4-Text at immutable revision cf8a174746cd14796c81ca2b54e035dc32e69bd8. It stores the upstream BF16 distilled transformer directly in mere.run's native module-key layout, together with the shared connector, video VAE, audio VAE/BWE vocoder, spatial upsampler, and duration head from Lightricks/LTX-2.5@dd53cc2.... The custom Gemma 4 language tower is MLX affine Q4/group-64; its LTX projection and tokenizer support remain BF16/raw. The complete selected payload is exactly 53,878,517,792 bytes. The repository is ungated, includes the governing LTX and Gemma terms and modification notice, and requires no separate component pull, local quantization, or post-install re-keying.

This model runs natively through Swift and MLX; no Python process or sidecar is dispatched. Use --quality final --output-mode audio-video for synchronized video and stereo audio.

video-ltx25-full-bf16

The complete official LTX 2.5 root is:

.../models/video-ltx25-full-bf16

mere.run model pull video-ltx25-full-bf16 --accept-model-license pulls Sawfwair/LTX-2.5-Full-BF16-MLX at immutable revision ac74d124f7211fc3cb8b32f418a08d8e71655c8d. It includes native-layout dev and distilled BF16 transformers, one shared connector, the official BF16 text encoder, distilled LoRA, diffusion video VAE, temporal x2 latent upsampler, and duration head. The complete selected payload is exactly 119,718,579,164 bytes. The repository is gated and needs no local optimization or source-transformer copy. This root supports full/dev and HQ pipelines, source-audio A2Vid, DFR, Retake, HDR/EXR, IC-LoRA reference video, Dub-It, and native text-to-audio.

The official DFR pixel-space spatial upscaler is a separately gated managed adapter: ltx25-pixel-spatial-upscaler-x2. Its repository gate and mere.run adapter pull ... --accept-license acknowledgement are independent of the main model gate.

See LTX 2.5 upstream parity for the exact pinned code release and pipeline matrix.

MiniMax-H3 FL2VA and Ref2VA

The native MiniMax-H3 implementation covers both released 33B dense partitions: FL2VA text/first/last-frame conditioning and Ref2VA ordered image/video/audio conditioning. It jointly denoises 24-channel video and 32-channel stereo-as-batch audio latents, then decodes 24 fps RGB video and 32 kHz stereo audio into one MP4.

The legacy explicit-pull FL2VA root is the flat Sawfwair/MiniMax-H3-FL2VA-MLX-4bit package pinned at immutable Hub commit e1244ad93d60c737c7e0f065a1c9372f3de7caf8. Every tensor in that package is derived directly from MiniMaxAI/MiniMax-H3@ec19cc6daf5d8add9417c18e86b6b58cc6c55027; converted or quantized third-party weights are not inputs. The transformer core is affine Q4/group-64 with dense precision islands, the exact 50-layer Qwen3-VL conditioner is affine Q8/group-64, the video VAE is FP16, and the audio VAE has its released weight normalization folded for the native runtime. The 14-file managed download is exactly 46,250,104,566 bytes and also carries the tokenizer, configuration, source manifest, conversion receipt, hashes, LICENSE, NOTICE, and MODIFICATIONS.md. Before Q4 packing, all 52 fused transformer QKV matrices are deinterleaved from MiniMax's released per-head row order into the global Q/K/V slabs consumed by the native runtime. Runtime auto-download is disabled. This compatibility ID remains installable, but Q4 did not meet the release visual-quality bar and is no longer recommended.

The maximum-fidelity video-minimax-h3-fl2va-bf16-mlx ID now resolves to the single-root Sawfwair/MiniMax-H3-FL2VA-MLX-BF16 compact artifact pinned at immutable Hub commit 4ce4b1d870f7b1b0c75672fd4f2867c1f5df7b5f. Its active 20.11B-parameter denoising core remains BF16 while the Q8 conditioner, FP16 video VAE, FP32 audio VAE, tokenizer, license, provenance, and cache pack live beside it. The verified runtime payload is exactly 77,094,088,403 bytes. The old full transformer overlay remains recoverable from pipenetwork/MiniMax-H3-MLX-bf16@1486555759eed9e3037edf29f9e055a0713bab2f, but is no longer a managed download source.

The production AdaLN tables in both compact artifacts are generated from the pinned official projection tensors by mere.run model optimize on MLX Metal. Their source-closure receipt includes exact matched 9- and 21-point real-media hashes. A streaming MLX 0.32.1 Metal refresh adds the exact LightX2V 9-point shifts-6/3 table and proves that the seven pre-existing tables retain their tensor closures. CUDA is still used to reproduce and quantize the large transformer cores, but CUDA-generated BF16 modulation tables are rejected because backend reduction order is not bit-identical to Apple Silicon.

video-minimax-h3-fl2va-8bit-mlx uses the sibling Sawfwair/MiniMax-H3-FL2VA-MLX-8bit artifact pinned at immutable Hub commit 86500cb6ebec22c006597e41840b26ef1099fdd7. Eligible core linears use MLX affine INT8/group-64; the conditioner, VAEs, tokenizer, cache tables, and official-source provenance remain the same as compact BF16. Its managed runtime payload is exactly 58,308,237,969 bytes. Q8 is a disk and memory option instead of a speed claim. Both compact packages are explicit-pull only and runtime auto-download remains disabled.

The explicit-pull Ref2VA root is the flat Sawfwair/MiniMax-H3-Ref2VA-MLX-8bit package pinned at immutable Hub commit 61dc387ef1a7166425cdacd63c2340598dcc364f. Its 14-file managed download is exactly 70,941,103,245 bytes and includes the complete runtime root, a source-bound 31-point AdaLN cache, source manifest, conversion receipt, hashes, license, notice, and modification disclosure. Runtime auto-download is disabled. Eight-bit is the supported Ref2VA floor because lower precision did not meet the visual quality bar.

The audited release converter accepts only minimax_h3_ref2va_int8_convrot.safetensors from Comfy-Org/MiniMax-H3@fd70b39279d1ae6eb214c903f53e1bec3af19a77, exactly 34,038,894,550 bytes with SHA-256 9eef934046a0671bc8a5daf87100705e1478419c574cfde70c50fbe6885f76a9. It validates each tensor's embedded ConvRot group size, reverses that regular-Hadamard basis, reproduces MLX affine INT8 group-64 packing, and emits a hashed receipt. The source uses group 256 for 200 transformer matrices and group 64 for 50 AdaLN matrices; these source groups are independent of MLX's output group size. The verified CPU conversion from the script's pinned PyTorch 2.7.1 toolchain is 36,024,412,656 bytes with SHA-256 234f22f69f8d40d6ed81cceed8259fa287f3c9417d40fba5274e3a7aa84e18a2. It is published as transformer.safetensors beside the exact FL conditioner, VAEs, tokenizer, notices, and a config.json whose partition is ref2va.

convert_minimax_h3_official_mlx.py creates the managed FL2VA packages in one audited pass from the official release. It preserves the active transformer as BF16 or directly quantizes eligible core linears to affine Q8/group-64 and always quantizes the conditioner to Q8 once. Before omitting the schedule-only AdaLN projections, timestep MLP, and reconstructed RoPE tensors, it evaluates source-bound exact tables for 5, 9, 12, 16, 21, and 31 points at video/audio shifts 12/3 plus the LightX2V 5- and 9-point shifts 6/3 schedules. The versioned pack index binds each geometry to its filename, byte count, SHA-256, and official transformer identity. Custom schedules interpolate from the densest table and are visibly disclosed as not bit-exact.

The model weights use the MiniMax-H3 Community License, not Apache-2.0. At the pinned official source revision ec19cc6daf5d8add9417c18e86b6b58cc6c55027, the license excludes use, distribution, and display in the United States, European Union, United Kingdom, and Republic of Korea and imposes notice, modification-disclosure, and safety obligations on downstream distribution. The CLI therefore requires explicit license acceptance for the managed FL pull, and conversion must be performed only where the model terms permit it. The upstream FL2VA and Ref2VA trees are about 144 GB each; the complete upstream repository is roughly 498 GB, so compile success is not artifact or generation proof.

The optional minimax-h3-turbo-4step runtime adapter is the EMA checkpoint minimax_h3_turbo_4step_ema_ckpt850.safetensors from larryvrh/MiniMax-H3-Turbo-Lora, pinned at immutable commit b604dd5fe25c4c747699f698a1e63f6c46d4a066. The catalog verifies its exact 779,849,816-byte length and SHA-256 5a6eeba171cf183020a4ad48774bb2968f29f8168afd6ec17a04987f3528b4ea. The adapter card declares Apache-2.0, but using it does not replace or relax the MiniMax-H3 Community License governing the required BF16 base model.

The managed minimax-h3-fasth3-vsa-datafree-4step adapter pins vsa-datafree/adapter_model.safetensors from FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA at immutable revision bcf40ca6f457ed66f8badf13514943e390205fca. The catalog verifies its exact 5,339,117,712-byte length and SHA-256 42dc502a2078f166c396a1fa75f29728d1844363652d345d5ef3e2b444ed6470. The adapter contains 362 LoRA pairs, 82 direct differences, and 50 VSA-H3 compression-gate weights. The runtime consumes the released fastvideo-lora-v2 format without producing a fused checkpoint.

The managed video-minimax-h3-fasth3-vsa-datafree-mlx package stores a premerged affine Q8/group-64 student in Sawfwair/MiniMax-H3-FastH3-VSA-DataFree-MLX-Q8 at immutable revision 6068ae3dafafb1e4b2afb29f3109745a16912e07. The package also includes the Q8 text encoder, both VAEs, tokenizer, Q8 compression gates, and source-bound AdaLN table. One explicit license-accepting pull installs every asset used by supported FastH3 inference. The runtime doesn't fetch a second model or adapter during generation.

FastH3 changes the base transformer's time and AdaLN projections. The preparation script range-reads the required tensors from FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree at immutable revision b65818d41939b5085451074fe8ca8b799f8d4921. The script verifies the transformer index SHA-256, evaluates the four released DMD points with MLX Metal, and stores the source-bound cache beside the compression gates. The offline merge tool verifies all 362 LoRA pairs, 82 direct differences, and 50 compression gates. It merges 208 inference linear targets with FP32 accumulation, rounds once to BF16, and encodes the result as MLX affine Q8/group-64. The runtime loads the premerged transformer and doesn't apply live LoRA deltas. The MiniMax-H3 Community License continues to govern the base and student weights.

Another managed H3 adapter, minimax-h3-lightx2v-4step, pins minimax_h3_fl2v_turbo_4step_v0.1.safetensors from lightx2v/Minimax-h3-Turbo at immutable commit b65e359c0d128b3c5e08e0f5bf2791b794378588. The catalog verifies its exact 1,383,677,888-byte length and SHA-256 5ff4a12c8b4599fec716e1b15a45e504e0d1129111896bdcde5ac4a15e395b29. The runtime consumes its 312 published PEFT pairs directly, applies the published alpha/rank scale, and projects its separate Q, K, and V deltas into the base transformer's global slabs without producing an expanded converted checkpoint. It fuses each scaled delta into the BF16 transformer once during checkpoint. Standard and QKV projections run as activation-space low-rank wrappers around dense or stock MLX affine linears; base weights are never fused or expanded. Its Apache-2.0 adapter license likewise does not replace the base model's MiniMax-H3 Community License.

The v1.0 managed adapters minimax-h3-lightx2v-8step-v1 and minimax-h3-lightx2v-4step-v1-768p pin the non-ComfyUI BF16 checkpoints at immutable repository revision e6346777701aa2b64d42ed058cdd71ae00e7cd52. Their exact sizes and SHA-256 digests are 1,383,677,768 bytes / e16ac20824d6e6649b193806f8fb095639bd9946c97b1bb84b4248eab1cc807f and 1,383,677,808 bytes / 1bdabc2e9fce20b1db563b96bcf6e46adcad4c1964f423676436bf266cc7416c. The runtime binds each filename to the upstream recipe so the 8-step release uses shifts 12/3 and alpha 8, while the 1344x768 four-step release uses shifts 6/3 and alpha 128.

The minimax-h3-lightx2v-8step-v1-768p adapter pins the non-ComfyUI BF16 checkpoint minimax_h3_fl2v_turbo_8step_v1.0_768p_bf16.safetensors at immutable repository revision 05ef678438e84933c406131b59abbf86919b3aac. The catalog verifies its exact 1,383,677,808-byte length and SHA-256 9b0efe3613b43a84e30febaa43af27432ea9d0711eac7bba904b2556b175f6d4. The checkpoint contains 312 BF16 PEFT pairs at rank 128 and declares alpha 8. mere.run selects eight model evaluations and video/audio shifts 6/3 for its 1344x768 recipe. The pinned compact model roots select the exact source-bound nine-point AdaLN table for this schedule; full BF16 source roots compute the same table from their AdaLN weights.

The Ref2VA-specific minimax-h3-lightx2v-ref2v-4step-v0.1 adapter pins the non-ComfyUI BF16 checkpoint minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors at immutable repository revision 5d1d4829fe614c1b93fcfd9cc7718e9ba71f73e1. The catalog verifies its exact 1,383,677,768-byte length and SHA-256 9e642fc8749c74f8da5e2382877ab5c7aa37b9a73b7fd0d6d457bd1b3cb1ae99. The checkpoint contains 312 BF16 PEFT pairs at rank 128. mere.run applies the published alpha 8 and video/audio shifts 12/3, expands the managed INT8 Ref2VA transformer to resident BF16, and fuses the adapter once before its four denoise evaluations. The adapter is Apache-2.0; the required base model remains governed by the MiniMax-H3 Community License.

video-wan22-ti2v-5b-mlx

The native Wan2.2 TI2V-5B model root is:

.../models/video-wan22-ti2v-5b-mlx

mere.run model pull video-wan22-ti2v-5b-mlx installs a pinned MLX conversion of Wan-AI/Wan2.2-TI2V-5B. The root contains the single 5B transformer, local UMT5 tokenizer and encoder, and the float32 48-channel Wan2.2 VAE required for image conditioning. The managed snapshot is pinned by revision; inference does not fetch tokenizer or model components at runtime.

The core runtime also exposes Wan2WorldSession for long-lived world generation. One session retains the models, prompt cache, and terminal-frame latent across transitions; callers receive MP4/PNG state artifacts and opaque state IDs rather than mutable MLX tensors. Camera controls are represented as XYZ translation and rotation so a DreamX-derived causal camera conditioner can replace the text-plus-first-frame mode without changing the session schema.

video-scail2-14b-mlx

The native SCAIL-2 model root is:

.../models/video-scail2-14b-mlx

The pinned Sawfwair/SCAIL-2-14B-MLX release package contains sharded BF16 transformer and UMT5 weights, FP16 OpenCLIP weights, the Wan 2.1 VAE, tokenizer, configuration, MIT license, model card, and a conversion receipt with immutable source and artifact hashes. The package is approximately 43 GiB.

The mere.run repository contains only the native Swift/MLX runtime. It does not ship a checkpoint converter or an upstream Python reference implementation. Install the immutable managed snapshot with:

mere.run model pull video-scail2-14b-mlx

An explicitly prepared compatible root can still be supplied with --model-root.

The optional scail2-lightx2v-4step adapter is not part of that model package or this repository. mere.run adapter pull downloads only wan2.1_i2v_lora_rank64_lightx2v_4step.safetensors from lightx2v/Wan2.1-Distill-Loras at immutable revision 27ae38da91014b947dd39cc3fa78b97cd7b386dd, verifies 739,472,104 bytes and SHA-256 8833bd4fd7c8eabebf0bc8ee5cfaf47f4f310ce116928a02c1adf8941dd4b0f1, and stores it outside git under the managed adapter directory. The upstream model card declares Apache-2.0. The Wan 2.2 adapters are deliberately rejected for SCAIL-2 because they target the incompatible Wan 2.2 MoE base.

video-cosmos3-edge-mlx

The native Cosmos3-Edge root is the complete official nvidia/Cosmos3-Edge snapshot pinned at revision 6f58f6b4c91288838e60b6bcb2cc45d997e961de:

.../models/video-cosmos3-edge-mlx

The approximately 9.2 GB snapshot includes the mixed understanding/generation transformer, Wan VAE, scheduler, generation tokenizer, reasoner tokenizer/configuration, packed SigLIP2 vision encoder, and multimodal projector. The Swift/MLX runtime loads these official safetensors directly; it does not invoke Python, PyTorch, or Diffusers during inference.

mere.run model pull video-cosmos3-edge-mlx
mere.run guide video-cosmos3

The checkpoint is governed by NVIDIA Open Model Development and Use License 1.1 (OpenMDW-1.1). Exercising the licensed rights constitutes acceptance; the public, ungated pull does not require mere.run's acceptance flag. The managed snapshot retains LICENSE.md, and runtime auto-download remains disabled.

Numerical parity fixtures are generated against NVIDIA's Cosmos framework commit ed8287fd7477113f8ac4f6b84290514d55cf0cdc; the VAE/scheduler reference is Diffusers v0.39.0 commit a3608b512ed7248499a44c61d954965ed9bdae4d.

video-dreamx-world-5b-ar-mlx

The native DreamX-World autoregressive checkpoint root is:

.../models/video-dreamx-world-5b-ar-mlx

The public CLI does not convert this local-only checkpoint. Supply a previously converted GD-ML/DreamX-World-5B root at the documented managed path or pass its directory to mere.run world serve --model. The runtime pairs it with video-wan22-ti2v-5b-mlx for tokenizer, text encoder, and VAE resources. The checkpoint provides learned camera conditioning, block-causal attention, persistent attention caches, and autoregressive forcing for long-lived local world sessions. It is never downloaded or converted automatically at runtime.

See Persistent world runtime for the server and request lifecycle.