Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. Try the Lab · All OptiQ quants · Docs

A 4-bit mixed-precision MLX quant of Qwen/Qwen3.8-27B, an image+text model. Per-layer bit-widths come from a KL-divergence sensitivity pass on a six-domain calibration mix (prose · reasoning · code · agent · tool-call · constraint-bearing instructions). Sensitive layers go to 8-bit; robust ones stay at 4-bit.

The vision tower is kept at bf16 in a sidecar, so this repo loads text-only under stock mlx-lm and image+text under OptiQ.

Quantization details

Property	Value
Predominant precision	4-bit
Layers at 8-bit (sensitive)	261
Layers at 4-bit (robust)	237
Total quantized layers	498
Group size	64
Vision tower	bf16, `optiq/optiq_vision.safetensors` (0.9 GB)
Calibration mix	six-domain mix (40 samples × 6 domains)
Reference for sensitivity	uniform-4-bit (bf16 at this size does not fit 36 GB)
On-disk size	19.0 GB

Qwen3.8-27B is a hybrid: three quarters of its 64 blocks use linear attention (five projections each) and the rest full attention, which is why the sweep scores 498 tensors rather than the ~450 a dense model of this size would have.

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

Usage

Load it with mlx-lm for text:

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Qwen3.8-27B-OptiQ-4bit")
response = generate(
    model, tokenizer,
    prompt="Explain quantum computing in simple terms.",
    max_tokens=200,
)

For image+text, mixed-precision KV-cache serving, sensitivity-aware LoRA fine-tuning, an OpenAI + Anthropic-compatible inference server, hot-swap mounted adapters and sandboxed Python execution, install mlx-optiq:

bash

pip install mlx-optiq
optiq serve --model mlx-community/Qwen3.8-27B-OptiQ-4bit

See the Qwen family guide on mlx-optiq.com for sampling defaults, training recipes, and family-specific caveats.

Benchmarks

Metric	Score
MMLU (5-shot, 1000)	84.1%
GSM8K (3-shot CoT)	96.1%
IFEval (prompt-level, strict)	77.1%
BFCL-V3 simple (AST)	93.5%
HumanEval (pass@1)	92.1%
HashHop (long-context retrieval)	85.0%
Capability Score (mean of 6)	87.98

18.9 GB on disk.

Every metric gets one equal vote, and disk size is reported next to the score as an honest second axis rather than folded into it. See the eval-framework writeup for the methodology.

Quantize your own

This quant was produced by mlx-optiq. Point it at any Hugging Face model to get the same sensitivity-aware mixed precision:

bash

pip install mlx-optiq
optiq convert Qwen/Qwen3.8-27B --target-bpw 5.0 --candidate-bits 4,8
optiq lab   # or drive the whole thing from the browser

mlx-community/Qwen3.8-27B-OptiQ-4bit

mlx-community/Qwen3.8-27B-OptiQ-4bit

Quantization details

Usage

Benchmarks

Links

Quantize your own

mlx-community/
Qwen3.8-27B-OptiQ-4bit