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, built on Google's quantization-aware-trained (QAT) Gemma-4 base. OptiQ's sensitivity-guided per-layer bit allocation is applied on top of weights that were trained to survive low-bit quantization, and it still beats a uniform 4-bit quant of the same QAT base by +1.37 Capability Score points.

This is a quant of google/gemma-4-12B-it-qat-q4_0-unquantized. 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.

Quantization details

Property	Value
Base	google/gemma-4-12B-it-qat-q4_0-unquantized (QAT)
Predominant precision	4-bit
Components at 8-bit (sensitive)	157
Components at 4-bit (robust)	171
Total quantized components	328
Achieved bits-per-weight	5.25
Group size	64
Reference for sensitivity	uniform 4-bit (streamed)
Calibration mix	six-domain mix
Vision	bf16 sidecar (`optiq_vision.safetensors`), image+text via optiq
Speculative drafter	`google/gemma-4-12B-it-qat-q4_0-unquantized-assistant` via `optiq serve --drafter`

Capability Score

Six-metric mean (MMLU, GSM8K, IFEval, BFCL, HumanEval, HashHop), scored against a uniform 4-bit quant of the same QAT base. That comparison isolates what the mixed-precision allocation adds, holding the base fixed.

Benchmark	This model (OptiQ, QAT base)	Uniform-4 (QAT base)	Delta
MMLU (5-shot, 1000)	52.5%	50.9%	+1.6
GSM8K (1000)	93.3%	93.1%	+0.2
IFEval (full, strict)	73.6%	72.3%	+1.3
BFCL-V3 simple (200)	72.0%	72.5%	-0.5
HumanEval (pass@1, 164)	91.5%	90.9%	+0.6
HashHop (long-context)	35.0%	30.0%	+5.0
Capability Score (mean)	69.64	68.27	+1.37

OptiQ adds +1.37 points over uniform 4-bit on this QAT base, consistent with the margin on the other QAT Gemma-4 sizes (E2B +2.09, E4B +1.19): the per-layer allocation keeps paying off even after QAT has made the weights more quantization-robust. The mixed quant is 5.25 bits-per-weight (about 8.3 GB on disk) versus 4.0 bits-per-weight (about 6.2 GB) for uniform 4-bit, with the extra budget spent on the layers that need it.

Usage

The 12B is the unified Gemma-4 (model_type: gemma4_unified), so it needs mlx-lm from main and import optiq (the unified text tower is not in the 0.31.3 PyPI release; the main build also reports 0.31.3, so install from git, not a version pin):

pip install -U mlx-optiq "mlx-lm @ git+https://github.com/ml-explore/mlx-lm.git"

python

import optiq  # registers the gemma4_unified model type
from mlx_lm import load, generate
model, tokenizer = load("mlx-community/gemma-4-12B-it-qat-OptiQ-4bit")
print(generate(model, tokenizer, "Explain mixed-precision quantization.", max_tokens=256))

Image+text input and the speculative drafter run through mlx-optiq:

bash

pip install mlx-optiq
optiq serve --model mlx-community/gemma-4-12B-it-qat-OptiQ-4bit \
            --drafter google/gemma-4-12B-it-qat-q4_0-unquantized-assistant

The language and image+text paths both run through optiq. The bf16 vision tower rides in optiq_vision.safetensors, which mlx-lm ignores (it globs model*.safetensors), so both paths work from one artifact.