UraionLabs/MiniCPM5-2B-oQ5e

Uraion Labs
_{Foundational systems research.}

MiniCPM5-2B — oQ5e

An MLX mixed-precision quantization of OpenBMB/MiniCPM5-2B for local inference on Apple Silicon, published by Uraion Labs.

MiniCPM5-2B-oQ5e is an MLX mixed-precision quantization of OpenBMB/MiniCPM5-2B optimized for local inference on Apple Silicon Macs. It was quantized and published by Uraion Labs using the oMLX oQe quantization workflow with importance-matrix guided sensitivity allocation. The checkpoint uses the standard MLX format and is directly compatible with both oMLX and mlx-lm.

About This Quantization

• Quantization Level: oQ5e
• Quantization Workflow: oMLX oQe mixed-precision quantization
• Base Weight Precision: 5-bit
• Mixed-Precision Profile: Mixed 5/6/8-bit precision (19 @ 6b, 6 @ 8b, lm_head @ 6b)
• Layer Overrides: 25 overrides: 19 layers boosted to 6-bit, 6 layers boosted to 8-bit, and lm_head preserved at 6-bit
• Output Head (lm_head) Precision: 6-bit
• Group Size: 64
• Quantization Mode: Affine (scale and bias per group)
• Non-Quantized Tensor Precision: BF16 (LayerNorm weights, embedding scales and biases)
• Calibration Dataset: oqe_code_multilingual (294 calibration samples)
• Model File Size: 1.67 GB (1,710.44 MB)
• Effective Bits Per Weight: Effective ~5.4 bits per weight with critical layers preserved at 6-bit and 8-bit.

Model Information

Property	Value
Original Model	OpenBMB/MiniCPM5-2B
Quantization Variant	oQ5e
Quantized By	Uraion Labs
Architecture	LlamaForCausalLM (llama)
Base Precision	5-bit
Mixed-Precision Allocation	Mixed 5/6/8-bit precision (19 @ 6b, 6 @ 8b, lm_head @ 6b)
Output Head (lm_head)	6-bit
Quantization Mode	Affine (group size 64)
Non-Quantized Precision	BF16
Total Parameters	2,516,756,480 (~2.52B)
Non-Embedding Parameters	1,981,982,720 (~1.98B)
Layers	42
Attention Configuration	GQA — 16 Q heads / 2 KV heads (head dim 128)
Context Length	131,072 tokens
Storage Format	MLX safetensors
Target Runtime	Apple Silicon macOS (omlx, mlx-lm)
Model Weight Size	1.67 GB (1,710.44 MB)
License	Apache-2.0

Upstream MiniCPM5-2B Highlights

MiniCPM5-2B is developed by OpenBMB as the 2B-scale model in the MiniCPM5 series:

• 2B-Class Open-Source State of the Art: Reaches top-tier performance among models in its size category (averaging 53.9 across OpenBMB's evaluation set) while remaining competitive with 3B and 4B class models.
• Native 131k Context: Supports up to 131,072 tokens context window out of the box for document synthesis, repo-wide code reasoning, and long conversation threads.
• Standard LLaMA Architecture: Uses standard LlamaForCausalLM with Grouped-Query Attention (GQA, 16 Q heads, 2 KV heads), enabling out-of-the-box execution across standard inference runtimes.
• High-Quality Training Datasets: Trained using OpenBMB's UltraData data curriculum, including Ultra-FineWeb, UltraX, UltraData-Code, UltraData-SFT, and UltraData-RL.
• Agentic and Tool-Use Capabilities: Engineered for high accuracy on function calling, structured output generation, and coding assistant workflows.

Quantization Family Navigation

Quant	Base Bits	Mixed-Precision Profile	lm_head	Model Size	Repository
oQ8e	8-bit	Uniform 8-bit	8-bit	2.49 GB	UraionLabs/MiniCPM5-2B-oQ8e
oQ6e	6-bit	Mixed 6/8-bit (28 layers @ 8-bit)	6-bit	1.95 GB	UraionLabs/MiniCPM5-2B-oQ6e
oQ5e	5-bit	Mixed 5/6/8-bit (19 @ 6b, 6 @ 8b)	6-bit	1.67 GB	UraionLabs/MiniCPM5-2B-oQ5e
oQ4e	4-bit	Mixed 4/5/6-bit (58 @ 5b, 9 @ 6b)	4-bit	1.38 GB	UraionLabs/MiniCPM5-2B-oQ4e
oQ3.5e	3-bit	Mixed 3/5/6-bit (29 @ 5b, 7 @ 6b)	6-bit	1.17 GB	UraionLabs/MiniCPM5-2B-oQ3.5e
oQ3e	3-bit	Mixed 3/5/6-bit (31 @ 5b, 7 @ 6b)	3-bit	1.08 GB	UraionLabs/MiniCPM5-2B-oQ3e
oQ2.7e	2-bit	Mixed 2/5/6/8-bit (23 @ 5b, 8 @ 6b, lm_head @ 8b)	8-bit	0.98 GB	UraionLabs/MiniCPM5-2B-oQ2.7e
oQ2e	2-bit	Mixed 2/5/6-bit (10 @ 5b, 6 @ 6b)	6-bit	0.88 GB	UraionLabs/MiniCPM5-2B-oQ2e

Quickstart

Using oMLX

oMLX provides a high-throughput runtime for MLX models on Apple Silicon:

# Install oMLX
pip install omlx

# Run inference
omlx run UraionLabs/MiniCPM5-2B-oQ5e --prompt "Explain quantum entanglement in simple terms."

Using mlx-lm

This quantization is fully compatible with Apple's standard mlx-lm library:

# Install mlx-lm
pip install mlx-lm

Python Example

from mlx_lm import load, generate

model, tokenizer = load("UraionLabs/MiniCPM5-2B-oQ5e")

prompt = "Explain quantum entanglement in simple terms."
messages = [{"role": "user", "content": prompt}]
prompt_formatted = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

response = generate(
    model,
    tokenizer,
    prompt=prompt_formatted,
    max_tokens=512,
    temp=1.0,
    top_p=0.95,
    verbose=True,
)

CLI Example

python -m mlx_lm.generate \
  --model UraionLabs/MiniCPM5-2B-oQ5e \
  --prompt "Explain quantum entanglement in simple terms." \
  --temp 1.0 \
  --top-p 0.95 \
  --max-tokens 512

Hugging Face Download

To download the repository files locally:

hf download UraionLabs/MiniCPM5-2B-oQ5e

Recommended Generation Settings

The following sampling parameters are officially recommended by OpenBMB for MiniCPM5-2B:

Parameter	Recommended Value	Note
temperature	1.0	Balanced diversity and reasoning coherence
top_p	0.95	Nucleus sampling threshold
do_sample	true	Stochastic sampling enabled

When performing deterministic evaluation or greedy decoding, set temperature=0.0 or do_sample=false.

Evaluation Results and Benchmark Context

Benchmark results below are reported by OpenBMB for the original MiniCPM5-2B release. This Uraion Labs quantization has not been independently benchmarked unless explicitly stated otherwise.

Capability / Benchmark	MiniCPM5-2B (Upstream)	Qwen3.5-4B	granite-4.2-3B	LFM2.5-2.6B
Average Score	53.9	51.1	42.7	33.2
LiveCodeBench v6	69.1	56.4	58.9	42.1
LCB-Pro 25Q2 (Easy)	68.0	58.3	54.6	30.9
AIME 2025	86.5	78.8	79.4	41.9
AIME 2026	86.5	82.7	83.5	45.2
MATH-500	94.6	99.0	97.0	89.6
MMLU-Pro	70.8	78.0	65.8	65.2
MMLU-Redux	84.7	88.7	78.9	80.0
IFBench	66.3	59.0	73.0	59.0
AA-LCR (Long Context)	59.0	46.1	35.8	5.3

Limitations and Disclaimer

This model generates text based on statistical language modeling patterns and may produce inaccurate, biased, or hallucinated statements. Responses regarding sensitive or specialized domains (such as law, medical treatment, or financial investment) must not be treated as professional advice.

This model is provided "AS IS", without warranty of any kind, express or implied. Users are responsible for verifying outputs, implementing safety guardrails, and complying with all applicable local regulations and acceptable use policies.

Original Model and Attribution

This repository contains a quantized derivative of OpenBMB/MiniCPM5-2B.

MiniCPM5-2B was developed and trained by OpenBMB. Model architecture, pre-training, instruction tuning, benchmarks, and capabilities described on this page originate from the upstream MiniCPM5-2B release.

Training-data metadata is inherited from the upstream OpenBMB/MiniCPM5-2B release. Uraion Labs performed quantization only and did not retrain this checkpoint.

Upstream Resources

• Model Card: openbmb/MiniCPM5-2B
• GitHub Repository: OpenBMB/MiniCPM
• MiniCPM Technical Report: arXiv:2506.07900
• UltraData Paper: arXiv:2602.09003
• Online Demo: MiniCPM5-2B-Demo
• OpenBMB MLX Deployment Guide: mlx.md

License

This repository and the underlying MiniCPM model weights are released under the Apache-2.0 License.

Citation

Please cite the original MiniCPM research when referencing this work:

@article{minicpm4,
  title={Minicpm4: Ultra-efficient llms on end devices},
  author={MiniCPM, Team},
  journal={arXiv preprint arXiv:2506.07900},
  year={2025}
}

_{Quantized and published by Uraion Labs. Foundational systems research.}