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 of GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking, a reasoning fine-tune of MiniCPM5-1B. Sensitive layers are kept at 8-bit and robust ones at 4-bit.

2.0 GB of bf16 weights become 874 MB, small enough to run comfortably on any Apple Silicon Mac.

Quantization details

Property	Value
Predominant precision	4-bit
Layers at 8-bit (sensitive)	67
Layers at 4-bit (robust)	102
Total quantized layers	169
Achieved bits per weight	5.805
Group size	64
Size on disk	874 MB, from a 2.0 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. At 1B parameters there is less redundancy to spend, so the optimizer keeps more layers at 8-bit and the weighted average lands higher than on a larger model.

How the bit-widths were chosen

The per-layer allocation is transferred from mlx-community/MiniCPM5-1B-OptiQ-4bit, where it was derived by a KL-divergence sensitivity sweep against the bf16 reference on a six-domain calibration mix.

This model is a fine-tune of openbmb/MiniCPM5-1B with an unchanged architecture, so all 169 quantizable layers map across exactly and the allocation lands at the same 5.805 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.

Usage

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-OptiQ-4bit")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is 17 + 25?"}],
    add_generation_prompt=True, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

This is a reasoning model: it thinks inside <think>...</think> before answering, so give it enough max_tokens to finish.

For an OpenAI-compatible server, mixed-precision KV-cache serving, and sensitivity-aware LoRA fine-tuning, install mlx-optiq:

bash

pip install mlx-optiq
optiq serve --model mlx-community/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking-OptiQ-4bit

Verification

Generation was exercised on the finished artifact before release: raw completion, chat-template reasoning (it correctly reasons to 42 for 17 + 25), and free prose.

The quantization was also checked numerically: dequantizing individual layers out of the artifact and comparing them against the bf16 checkpoint gives 0.7% mean relative error on the 8-bit layers and 9.8% 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 MiniCPM5-1B OptiQ card.

Quantization does not change the behaviour or alignment of the base model. Use it under the same terms as the original.