A 4-bit mixed-precision MLX quant of migtissera/Tess-4-27B, an agentic reasoning model built on Qwen3.6-27B. Sensitive layers are kept at 8-bit and robust ones at 4-bit.

52 GB of bf16 weights become 19 GB, which fits a 24 GB Mac.

Image input works. The vision tower is kept at bf16 in a sidecar, so this quant takes images as well as text.

We follow the same naming convention llama.cpp uses for Q4_K_M and similar mixed-precision quants: the "4-bit" label is the predominant precision, not the weighted average.

The per-layer allocation is transferred from mlx-community/Qwen3.6-27B-OptiQ-4bit, where it was derived by a KL-divergence sensitivity sweep against the bf16 reference on a six-domain calibration mix.

Tess-4-27B is a fine-tune of Qwen/Qwen3.6-27B and its architecture is unchanged (every field of the text config matches), so all 496 quantizable layers map across exactly and the allocation lands at the same 4.769 bits per weight when recomputed against Tess's own tensors.

These are measured bit-widths, not a static rule-of-thumb recipe. But they were measured on the base model, not on this fine-tune. Fine-tuning shifts weights, so Tess's own per-layer sensitivities could differ somewhat from the base's. Which layers are fragile is mostly a property of the architecture, so the transfer is sound, but it is a transfer and you should know that.

Only the language tower is quantized. The vision tower stays at bf16, which is how every OptiQ VLM ships.

The sidecars live in an optiq/ subfolder, so a stock *.safetensors glob ignores them and mlx-lm sees a clean language model.

This is a reasoning model: it thinks inside <think>...</think> before answering, so give it enough max_tokens to finish.

Image input needs mlx-optiq, which loads the bf16 vision sidecar and feeds the merged embeddings to the quantized language tower:

Or serve it over an OpenAI-compatible endpoint that accepts image content parts:

The base ships a Multi-Token Prediction head, bundled here as optiq/mtp.safetensors:

Text, arithmetic reasoning, and image understanding were all exercised on the finished artifact before release.

The quantization was also checked numerically: dequantizing individual layers out of the artifact and comparing them against the bf16 checkpoint gives 0.7% mean relative error on the 8-bit layers and 9.8% on the 4-bit layers, which is what each bit-width should cost.

No task benchmarks were run on this quant; for measured quality numbers on the base architecture, see the Qwen3.6-27B OptiQ card.

Quantization does not change the behaviour or alignment of the base model. Use it under the same terms as the original.