Recursive Flint

Recursive Flint is a sparse mixture-of-experts chat model with tool calling and visible reasoning. mere.run runs it natively in Swift/MLX from either the published MLX INT4 export or the original BF16 safetensors, with no conversion step. Use INT4 on a 48 GB or larger Mac. Use BF16 when you have 128 GB and want the original weights.

Install

Both checkpoints are explicit installs; inference commands don't download them automatically.

Model ID	Weights	Download	Unified memory (minimum / recommended)
text-chat-flint-4bit	Published INT4 export	About 22.6 GB	48 GB / 64 GB
text-chat-flint-bf16	Original BF16	About 79.6 GB	128 GB / 128 GB

The memory values are catalog admission estimates, not measurements of this runtime's performance on the released checkpoints.

mere.run model pull text-chat-flint-4bit

To use a snapshot you already have, pass its directory to --model, or pass the matching managed model ID with --model-root. The directory's config.json must set model_type to flint. NVFP4 and GGUF checkpoints use different tensor formats, and this runtime rejects them.

Chat

The following command sets sampling and the output budget explicitly:

mere.run text chat --model text-chat-flint-4bit \
    --prompt "Write a Swift function that merges overlapping intervals." \
    --temperature 1 --top-p 0.95 --top-k 20 --max-tokens 4096

Serve the local API

After installation, start the native serving engine:

mere.run api serve --engine text-chat-flint --model text-chat-flint-4bit

The API accepts tools, stop sequences, seeds, penalties, and log probabilities. A single response can return up to 32,768 tokens.

Behavior

• Context. Chat defaults to an 8,192-token working context. Prompt and output share the configured limit, up to the checkpoint's 262,144-token maximum. An oversized prompt fails instead of truncating history.
• Cache. Sliding-window layers keep the last 512 tokens. Full-attention cache memory grows with the context.
• Reasoning. The model controls its own reasoning. Display options hide or show it but don't turn it off. Assistant history keeps reasoning between turns.
• Tools. Tool messages use the checkpoint's chat template. The runtime reads tool calls only outside reasoning blocks and applies the shared tool allowlist and parallel-call policy.

Limits

This runtime doesn't implement grammar-constrained JSON, strict tool-schema decoding, LoRA, speculative decoding, continuous batching, quantized KV caches, or NVFP4 loading. Publisher performance parity and long-context quality are outside the validation described on this page.

Validation

The local tests compare all logits of a deterministic three-layer synthetic model with an independent scalar reference. They cover dense and shared experts, sigmoid routing with selection-only probability bias, learned attention sinks, RoPE placement, and sliding-cache chunk boundaries. The fixture generator uses only the Python standard library and is test tooling; inference uses Swift/MLX.

On an M5 Max with 128 GiB of unified memory, both released checkpoints pass native Metal arithmetic, a typed calculate tool round trip, and retrieval of an exact access code from a prompt longer than 1,024 tokens. The BF16 run uses the original weights, and all fifteen shard hashes match the pinned publisher metadata. The INT4 run uses the published weights and the production loader's pinned artifact verification. Both runs use production precision defaults.

A five-token probe ([1, 33, 204, 1024, 42]) compares all 750,720 logits with an independent PyTorch 2.10 CPU implementation, which was first checked against the scalar fixture. For each checkpoint, native full-sequence prefill and chunk sizes 1, 3, and 5 agree with that reference on all five next-token choices. The logits differ by the following amounts:

Checkpoint	Full-sequence maximum absolute difference	RMS difference
Original BF16	3.6158	0.6936
Published INT4	1.3944	0.3199

These results establish bounded greedy-choice agreement. They don't establish numerical equivalence, broad model quality, or the declared context limit.

To repeat the installed functional regression, point it at either checkpoint:

MERERUN_TEST_MLX_DEVICE=gpu \
MERERUN_FLINT_CHECKPOINT_ROOT=/path/to/flint-checkpoint \
    swift test --filter FlintInstalledQualificationTests

Linux preparation includes the CUDA quantized-matrix synchronization and BF16 accumulation corrections. Their patching tests cover source drift, repeat application, and SwiftPM's read-only checkout files. CUDA execution on the refreshed dependency pin needs separate qualification.

Sources and license

The managed checkpoints are pinned to these immutable revisions:

Model ID	Source	Revision
text-chat-flint-4bit	RecursiveSI/flint-int4	3b93f9053b530488405ae9fca51d4900d6ce21f0
text-chat-flint-bf16	RecursiveSI/flint	2a247be0c0aec895ccb0ac7ca92229ec6df37288

Both checkpoints are published under Apache-2.0. Architecture references include the BF16 checkpoint's SGLang implementation and Recursive's inference engine. The INT4 loader verifies the publisher's SHA-256 values before it loads the packed tensors, and it keeps the export's INT8 embeddings and INT4 projections without requantization. The BF16 loader checks shard ownership, complete tensor coverage, shapes, and floating-point types before use.

For other text models, see the text runtime.