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 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/Qwen3.5-4B. 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

Property	Value
Predominant precision	4-bit
Layers at 8-bit (sensitive)	75
Layers at 4-bit (robust)	173
Total quantized layers	248
Group size	64
Calibration mix	six-domain mix (40 samples × 6 domains)
Reference for sensitivity	bf16 (auto-resolved; falls back to uniform-4-bit if bf16 doesn't fit)
Bundled MTP head	`mtp.safetensors` (4-bit projections, BF16 norms), enables 1.4× decode via `optiq serve --mtp`

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. 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:

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Qwen3.5-4B-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:

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:

optiq serve --model mlx-community/Qwen3.5-4B-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:

Metric	OptiQ	Uniform 4-bit	Δ
MMLU (5-shot, 1000 samples)	69.9%	68.7%	+1.1
GSM8K (1000 samples, 3-shot CoT)	80.5%	78.8%	+1.7
IFEval (full set, strict)	69.1%	68.4%	+0.7
BFCL-V3 simple (200 calls)	90.0%	85.0%	+5.0
HumanEval (164 problems, pass@1)	78.0%	76.2%	+1.8
HashHop (long-context retrieval)	25.0%	24.0%	+1.0
Capability Score (mean of 6)	68.76	66.86	+1.90
KL vs bf16 reference (mean / p95)	0.1224 / 0.5692	,	,
On-disk size	3.0 GB	2.8 GB	+0.2

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.

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).

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