Native Kolibri-1

Kolibri-1 runs through native Swift/MLX computation. It uses 50 transformer layers, 384 routed experts per layer, six selected experts per token, and one shared expert. Four sliding attention layers precede each full attention layer. Sliding layers apply RoPE; full layers use no positional encoding.

The checkpoint format supports a BF16 reference, an 8-bit variant, and explicit mixed quantization policies. Mixed 2-bit keeps routed experts at 2-bit with 128-value affine groups, attention and shared experts at 8-bit, and embeddings, norms, routers, and output-head weights at their original precision. Routing and the output head compute in FP32.

Download a measured artifact

The selected mixed 2-bit artifact and Q8 control are public. Both require mere.run v0.62.0 or later with the native Kolibri runtime. Use the immutable release revisions to reproduce the measured weights:

hf download Sawfwair/Kolibri-1-MLX-Mixed-2bit \
  --revision 73be205486dd71273a30ab9ce3441e548f0a33f4 \
  --local-dir ./Kolibri-1-MLX-Mixed-2bit

hf download Sawfwair/Kolibri-1-MLX-8bit \
  --revision 792bfb73b9ab915724ccfd58916587330a642c16 \
  --local-dir ./Kolibri-1-MLX-8bit

The publication receipt records all 64 bundle file hashes and sizes for each release. Managed download registration remains separate; pass the explicit checkpoint directory today.

Run a converted checkpoint

swift run mere.run text chat \
  --model /path/to/Kolibri-1-MLX-mixed2 \
  --prompt "Erkläre, wie ein Regenbogen entsteht." \
  --context-size 8192 --max-tokens 256

swift run mere.run api serve \
  --engine text-chat-kolibri --model /path/to/Kolibri-1-MLX-mixed2 \
  --context-size 8192

Use the converted checkpoint directory itself. The upstream FP8 and BF16 repositories use a different tensor layout and must be converted first.

The runtime uses the checkpoint's chat template and tokenizer, supports text, reasoning, tool messages, seeded sampling, stop sequences, and opt-in token logprobs. LoRA, constrained JSON, KV quantization, prefix reuse, and continuous batching are not implemented for this family. Native chat defaults to an 8,192-token context budget; set it explicitly when serving through the API, which otherwise uses the server's shared default. The runtime rejects oversized prompts; it does not silently truncate them. The architecture's 262,144-token limit is not a memory-fit claim.

Conversion and quality measurements

The remote conversion tools pin Aleph-Alpha/Kolibri-1-BF16 at 7a8f290e7858825c3cf5e4c447ba68345de9f1d3, verify source hashes, stack expert banks in numeric order, and record every output shard's checksum. Keep the 156 GB source and BF16 reference on a sufficiently large remote worker.

On October 6, 2026, Hugging Face returns HTTP 404 (Invalid rev id) for that recorded upstream revision. The public converted release revisions above still resolve. Reproducing conversion requires a preserved original source snapshot with the recorded checksums; a replacement upstream revision needs separate validation.

python scripts/model-conversion/convert_kolibri_mlx.py \
  --source /workspace/kolibri/source --output /workspace/kolibri/artifacts \
  --profiles reference mixed2 mixed2-down3 q8

python scripts/reference-parity/prepare_kolibri_suite.py \
  --source /workspace/kolibri/source --output /workspace/kolibri/suite

mere.run model benchmark kolibri-logprobs \
  --model-root /workspace/kolibri/artifacts/Kolibri-1-MLX-reference \
  --suite /workspace/kolibri/suite/suite.json \
  --output /workspace/kolibri/results/reference \
  --calibration-output /workspace/kolibri/results/input-moments.safetensors

mere.run model benchmark kolibri-logprobs \
  --model-root /workspace/kolibri/artifacts/Kolibri-1-MLX-mixed2 \
  --suite /workspace/kolibri/suite/suite.json \
  --output /workspace/kolibri/results/mixed2

python scripts/reference-parity/compare_kolibri_logprobs.py \
  --reference /workspace/kolibri/results/reference \
  --candidates /workspace/kolibri/results/mixed2 \
  --output /workspace/kolibri/results/comparison.json

For an activation-weighted Q2 candidate, use the reference's calibration statistics and retain the separate scoring receipt:

python scripts/model-conversion/refit_kolibri_mixed2.py \
  --reference /workspace/kolibri/artifacts/Kolibri-1-MLX-reference \
  --baseline /workspace/kolibri/artifacts/Kolibri-1-MLX-mixed2 \
  --moments /workspace/kolibri/results/input-moments.safetensors \
  --reference-results /workspace/kolibri/results/reference \
  --output /workspace/kolibri/artifacts/Kolibri-1-MLX-mixed2-calibrated

Every comparison uses identical token IDs, prompt boundaries, and teacher-forced continuations. It reports full-vocabulary KL divergence, next-token agreement, changes in target-token log probabilities, and perplexity ratios by case and language. Only cases explicitly labeled calibration contribute input second moments. Candidate selection uses the calibration split; heldout cases report quality without changing the fit. The included small English/German suite is a diagnostic, not a standardized benchmark or a general capability guarantee. English and German must also pass their diagnostic gates separately. Select between the baseline and fitted Q2 candidates; report the Q8 and 2/3-bit variants as controls rather than allowing their larger weight budgets to replace that Q2 selection.

The independent PyTorch synthetic fixture checks attention, routing, shared experts, norms, and chunked cache semantics on CPU and Apple GPU. The original BF16 PyTorch operator reference provides an additional checkpoint cross-check; it does not certify bitwise parity with official vLLM fused kernels.

The October 4 full-checkpoint diagnostic records the Runpod results and selects baseline mixed 2-bit by calibration KL. Its 24.66 GB tensor storage is a candidate for a 36 GB machine; the 83.65 GB Q8 control is a candidate for 128 GB. Actual Apple full-checkpoint memory fit and throughput still require validation.