CoolFace
Modelpublic

mlx-community/Qwen3.5-0.8B-OptiQ-4bit

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
23likes1.5kdownloads
Model Card

mlx-community/Qwen3.5-0.8B-OptiQ-4bit

Built with [mlx-optiq](https://mlx-optiq.com), 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 produced by mlx-optiq, the sensitivity-aware quantization toolkit for Apple Silicon. Beats stock uniform 4-bit on every benchmark in the six-metric Capability Score.

A 4-bit mixed-precision MLX quant of Qwen/Qwen3.5-0.8B. 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 on-disk size is within ~5 % of a stock uniform 4-bit MLX quant.

Quantization details

PropertyValue
Predominant precision4-bit
Layers at 8-bit (sensitive)56
Layers at 4-bit (robust)130
Total quantized layers186
Group size64
Calibration mixsix-domain mix (40 samples × 6 domains)
Reference for sensitivitybf16 (auto-resolved; falls back to uniform-4-bit if bf16 doesn't fit)
Bundled MTP headmtp.safetensors (4-bit projections, BF16 norms), enables 1.4× decode via optiq serve --mtp

We follow the same naming convention llama.cpp uses for Q4KM and similar mixed-precision quants: the "4-bit" label is for the predominant precision, not the weighted average. The mixed allocation is what lets this build beat stock uniform-4-bit on every benchmark below at the same disk size.

Usage

Load it with mlx-lm and use it as usual:

bash
pip install mlx-lm
python
from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Qwen3.5-0.8B-OptiQ-4bit")
response = generate(
    model, tokenizer,
    prompt="Explain quantum computing in simple terms.",
    max_tokens=200,
)

For more (mixed-precision KV-cache serving, sensitivity-aware LoRA fine-tuning, OpenAI + Anthropic-compatible inference server, hot-swap mounted adapters, sandboxed Python execution for agent workflows), install `mlx-optiq`:

bash
pip install mlx-optiq

Speculative decoding (MTP)

This quant ships with a bundled Multi-Token Prediction head as mtp.safetensors. Enable it for ~1.4× faster decode:

bash
optiq serve --model mlx-community/Qwen3.5-0.8B-OptiQ-4bit --mtp

Acceptance rate stays ~70% at depth 2 (the empirical sweet spot for Qwen3.5).

See the Qwen3.5 family guide on mlx-optiq.com for sampling defaults, training recipes, and family-specific caveats.

Benchmarks

Six-metric Capability Score (mean of MMLU + GSM8K + IFEval + BFCL + HumanEval + HashHop). Apples-to-apples comparison against stock uniform 4-bit:

MetricOptiQUniform 4-bitΔ
MMLU (5-shot, 1000 samples)51.1%48.5%+2.6
GSM8K (1000 samples, 3-shot CoT)37.3%31.8%+5.5
IFEval (full set, strict)55.6%49.5%+6.1
BFCL-V3 simple (200 calls)55.5%35.0%+20.5
HumanEval (164 problems, pass@1)25.0%20.1%+4.9
HashHop (long-context retrieval)6.0%13.0%-7.0
Capability Score (mean of 6)38.4232.98+5.44
KL vs bf16 reference (mean / p95)0.1060 / 0.3478,,
On-disk size0.6 GB0.6 GB+0.0

Every metric gets one equal vote. Disk size is reported next to the score as an honest second axis instead of being folded into the score. See the eval-framework writeup for the full methodology.

Links

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 <hf-model-id> --target-bpw 5.0 --candidate-bits 4,8
optiq lab   # full local workbench: chat, compare, quantize, fine-tune

License

Apache 2.0 (inherits from base model).