Modelpublic

tinyopsec/minicpm5-2b-distilled-reasoning-GGUF

sourceHugging Faceapache-2.0updated 10d agoView on Hugging Face

0likes1.2kdownloads

Model Card

minicpm5-2b-distilled-reasoning GGUF

GGUF quantizations of jigs97022/minicpm5-2b-distilled-reasoning.

A 2B parameter reasoning model distilled from three frontier models — Qwen3.8-Max, GLM-5.2, and Kimi K3 — onto the efficient openbmb/MiniCPM5-2B architecture. Trained via QLoRA on 10,000 quality-filtered reasoning traces covering math, code, logic puzzles, and structured problem-solving.

Quantization Table

File	Bits	Size	Use Case
`model_f16.gguf`	16	~5.0 GB	Maximum quality, reference
`model_q8_0.gguf`	8	~2.7 GB	Best quality / size tradeoff
`model_q6_k.gguf`	6	~2.1 GB	High quality
`model_q5_k_m.gguf`	5	~1.8 GB	Recommended
`model_q5_k_s.gguf`	5	~1.7 GB	Slightly smaller Q5
`model_q4_k_m.gguf`	4	~1.5 GB	Good balance
`model_q4_k_s.gguf`	4	~1.4 GB	Smaller Q4
`model_q3_k_l.gguf`	3	~1.2 GB	Low RAM, large variant
`model_q3_k_m.gguf`	3	~1.1 GB	Low RAM
`model_q3_k_s.gguf`	3	~1.0 GB	Minimum RAM Q3
`model_q2_k.gguf`	2	~0.8 GB	Extreme compression

VRAM Requirements

Quant	Min VRAM
F16	8 GB
Q8_0	4 GB
Q4KM	2 GB
Q2_K	1.5 GB

Training Details

Parameter	Value
Base Model	openbmb/MiniCPM5-2B
Teachers	Qwen3.8-Max, GLM-5.2, Kimi K3
Dataset	r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation
Train Samples	10,000
Method	QLoRA (r=64, alpha=32)
Max Context	4,096 tokens
Hardware	Kaggle T4 x2

Usage

llama.cpp

bash

./llama-cli -m model_q4_k_m.gguf \
  --temp 0.7 \
  -n 1024 \
  -p "If 3x + 7 = 22, what is x? Show step-by-step reasoning."

llama-cpp-python

python

from llama_cpp import Llama

llm = Llama(model_path="model_q4_k_m.gguf", n_ctx=4096)
output = llm(
    "Solve step by step: If 3x + 7 = 22, what is x?",
    max_tokens=1024,
    temperature=0.7,
)
print(output["choices"][0]["text"])

LM Studio

Search tinyopsec/minicpm5-2b-distilled-reasoning-GGUF in the model browser.

Ollama