mlx-community/MiniCPM-V-4.7-1B-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. All OptiQ quants · Docs

MiniCPM-V 4.7 1B for Apple Silicon: an OptiQ 4-bit language tower, with the vision tower kept at bf16.

This is the OptiQ quant of OpenBMB's MiniCPM-V 4.7 1B, an image-text model with a SigLIP NaViT vision tower and a dense Qwen3.5 language tower. OptiQ measured each layer's sensitivity, kept 64 of the 187 layers at 8-bit and put the other 123 at 4-bit. That raises the Capability Score from 48.17 for uniform 4-bit to 52.48, with the biggest gains on HumanEval (50.6 vs 32.9) and GSM8K (69.8 vs 65.1). Function calling is the one loss: BFCL drops from 48.5 to 41.5, mostly on calls the model doesn't make.

It works on image and text

Both ran through this quantized model with optiq serve, MLX only:

Image (the words "OPTIQ 42" on a white background, "What text is in the image?"):

OPTIQ 42

Text ("What is the capital of France? One word."):

Paris

What it is

Property	Value
Base	openbmb/MiniCPM-V-4.7-1B
Language tower	Qwen3.5, dense, 24 layers, hybrid linear/full attention, 262k context
Method	OptiQ mixed-precision, sensitivity-driven
Allocation	123 layers at 4-bit, 64 at 8-bit
Vision tower	SigLIP NaViT, bf16, in optiq/optiq_vision.safetensors
On disk	0.80 GB language + 1.10 GB vision
KV cache	kv_config.json, mixed 4/8-bit on the full-attention layers

Following llama.cpp's naming for mixed quants, the "4bit" label denotes the family, not the weighted average.

Capability Score

Six-metric mean (the standard OptiQ text eval), against a uniform 4-bit quant of the same base.

Metric	Uniform 4-bit	OptiQ
MMLU (5-shot, 969 samples)	54.0%	55.3%
GSM8K (1000 samples)	65.1%	69.8%
IFEval (full set, strict)	71.5%	74.7%
BFCL-V3 simple (200 calls)	48.5%	41.5%
HumanEval (164 problems, pass@1)	32.9%	50.6%
HashHop (long-context retrieval)	17.0%	23.0%
Capability Score (mean of 6)	48.17	52.48

Run it

MiniCPM-V 4.7 needs OptiQ 0.5.25 or later. Stock mlx-lm does not know the minicpmv4_7 architecture, so import optiq registers it, and OptiQ loads the vision sidecar.

pip install "mlx-optiq>=0.5.25"

For image input, serve it with an OpenAI- and Anthropic-compatible endpoint:

optiq serve --model mlx-community/MiniCPM-V-4.7-1B-OptiQ-4bit

For the measured mixed-precision KV cache, fetch kv_config.json from this repo and pass it in:

hf download mlx-community/MiniCPM-V-4.7-1B-OptiQ-4bit kv_config.json --local-dir .
optiq serve --model mlx-community/MiniCPM-V-4.7-1B-OptiQ-4bit --kv-config kv_config.json

Then send an image as image_url content. Text-only generation also loads directly:

import optiq  # registers minicpmv4_7 and the vision sidecar
from mlx_lm import load, generate

model, tok = load("mlx-community/MiniCPM-V-4.7-1B-OptiQ-4bit")
msgs = [{"role": "user", "content": "Explain why the sky is blue."}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
print(generate(model, tok, prompt=prompt, max_tokens=512))

Links

• Project website: mlx-optiq.com
• Family guide: mlx-optiq.com/docs/minicpm
• All OptiQ quants: mlx-optiq.com/models
• Base model: openbmb/MiniCPM-V-4.7-1B