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
22 GB of bf16 weights become 8.4 GB, which fits comfortably on a 16 GB Mac.
Image input works. Gemma-4-12B is the encoder-free gemma4_unified variant, and its multimodal embedder is kept at bf16 in a sidecar.
Quantization details
Property
Value
Predominant precision
4-bit
Layers at 8-bit (sensitive)
156
Layers at 4-bit (robust)
172
Total quantized layers
328
Achieved bits per weight
5.216
Group size
64
Multimodal embedder
bf16, in optiq/optiq_vision.safetensors
Size on disk
8.4 GB, from a 22 GB bf16 base
We follow the same naming convention llama.cpp uses for Q4_K_M and similar mixed-precision quants: the "4-bit" label is the predominant precision, not the weighted average.
This model is a fine-tune of google/gemma-4-12B-it with an unchanged architecture (every shape field of the text config matches), so all 328 quantizable layers map across exactly and the allocation lands at the same 5.216 bits per weight when recomputed against this model's own tensors.
These are measured bit-widths, not a static rule-of-thumb recipe. But they were measured on the base model, not on this fine-tune. Fine-tuning shifts weights, so this model's own per-layer sensitivities could differ somewhat from the base's. Which layers are fragile is mostly a property of the architecture, so the transfer is sound, but it is a transfer and you should know that.
Only the language tower is quantized. The multimodal embedder stays at bf16, which is how every OptiQ VLM ships.
Usage
gemma4_unified needs OptiQ's model registration, so load it through mlx-optiq:
pip install mlx-optiq
python
1import optiq # registers the gemma4_unified architecture with mlx-lm2from mlx_lm import load, generate
34model, tokenizer = load("mlx-community/gemma-4-12B-coder-fable5-composer2.5-v1-OptiQ-4bit")5prompt = tokenizer.apply_chat_template(6[{"role":"user","content":"Write a Python function to merge two sorted lists."}],7 add_generation_prompt=True, tokenize=False)8print(generate(model, tokenizer, prompt=prompt, max_tokens=512))
This model reasons before answering (it emits a <|channel>thought block), so give it enough max_tokens to finish.
Verification
Generation was exercised on the finished artifact before release: factual recall, arithmetic, and code generation (it correctly writes is_palindrome, reasoning about edge cases such as the empty string).
The quantization was also checked numerically: dequantizing individual layers out of the artifact and comparing them against the bf16 checkpoint gives 0.6-0.7% mean relative error on the 8-bit layers and 11-13% on the 4-bit layers, which is what each bit-width should cost.
No task benchmarks were run on this quant; for measured quality numbers on the base architecture, see the gemma-4-12B-it OptiQ card.
Quantization does not change the behaviour or alignment of the base model. Use it under the same terms as the original.
