gemma-4-26B-A4B-it-qat-OptiQ-4bit

mlx-community/gemma-4-26B-A4B-it-qat-OptiQ-4bit#

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. This is the sparse-MoE member of the family: 26B total parameters with 128 routed experts (about 4B active per token). OptiQ's sensitivity-guided per-layer bit allocation is applied on top of weights already trained to survive low-bit quantization, and it still beats a uniform 4-bit quant of the same QAT base on the Capability Score.

This is a quant of google/gemma-4-26B-A4B-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). On this MoE the sensitivity pass identifies the routed experts as where precision matters most, so the allocation puts 8-bit on the experts that need it (42 of 90 expert tensors) plus the sensitive attention and router layers, and keeps the robust ones at 4-bit.

Quantization details#

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.

The gain concentrates in the reasoning-heavy benchmarks (MMLU +1.6, GSM8K +1.1, IFEval +0.5), which is where the routed experts carry the most signal. On this sparse-MoE base, where most of the parameters live in 128 experts, the precision has to reach the experts to move capability, and the per-layer allocation puts it there. The mixed quant is 6.01 bits-per-weight (about 19 GB on disk) versus 4.0 bits-per-weight (about 14.5 GB) for uniform 4-bit.

Usage#

This is a Gemma-4 MoE (model_type: gemma4, gemma4_text experts), so it needs mlx-lm from main and import optiq (the MoE 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"

import optiq  # registers the OptiQ model paths
from mlx_lm import load, generate
model, tokenizer = load("mlx-community/gemma-4-26B-A4B-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:

pip install mlx-optiq
optiq serve --model mlx-community/gemma-4-26B-A4B-it-qat-OptiQ-4bit \
            --drafter google/gemma-4-26B-A4B-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.