Instructions to use jigs97022/minicpm5-2b-distilled-reasoning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jigs97022/minicpm5-2b-distilled-reasoning with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="jigs97022/minicpm5-2b-distilled-reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("jigs97022/minicpm5-2b-distilled-reasoning")
model = AutoModelForCausalLM.from_pretrained("jigs97022/minicpm5-2b-distilled-reasoning", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jigs97022/minicpm5-2b-distilled-reasoning with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jigs97022/minicpm5-2b-distilled-reasoning"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jigs97022/minicpm5-2b-distilled-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/jigs97022/minicpm5-2b-distilled-reasoning

SGLang

How to use jigs97022/minicpm5-2b-distilled-reasoning with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "jigs97022/minicpm5-2b-distilled-reasoning" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jigs97022/minicpm5-2b-distilled-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "jigs97022/minicpm5-2b-distilled-reasoning" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jigs97022/minicpm5-2b-distilled-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use jigs97022/minicpm5-2b-distilled-reasoning with Docker Model Runner:
```
docker model run hf.co/jigs97022/minicpm5-2b-distilled-reasoning
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

MiniCPM5-2B Distilled Reasoning

A high-performance 2B parameter reasoning model distilled from three frontier models (Qwen3.8-Max, GLM-5.2, Kimi K3) onto the efficient MiniCPM5-2B architecture. Trained on quality-filtered reasoning traces covering math, code, logic puzzles, and structured problem-solving.

Quick Start

Transformers / Unsloth

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained( "jigs97022/minicpm5-2b-distilled-reasoning", max_seq_length=4096, load_in_4bit=True, )

messages = [{"role": "user", "content": "If 3x + 7 = 22, what is x? Show step-by-step reasoning."}] inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda") outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7) print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Ollama (Local CPU/GPU)

ollama run hf.co/jigs97022/minicpm5-2b-distilled-reasoning-gguf

vLLM (API Server)

vllm serve jigs97022/minicpm5-2b-distilled-reasoning
--dtype half
--max-model-len 4096
--gpu-memory-utilization 0.8

Key Features

Multi-Teacher Distillation: Combines reasoning patterns from Qwen3.8-Max, GLM-5.2, and Kimi K3
Structured Reasoning: Explicitly trained to use tags for chain-of-thought
Efficient Inference: 2B parameters with 4-bit quantization runs at ~1.5GB VRAM
Native MiniCPM5 Template: Preserves original chat formatting for zero-shot compatibility
Broad Domain Coverage: Math, code generation, logical puzzles, tool-use reasoning

Training Details

Parameter	Value
Base Model	openbmb/MiniCPM5-2B
Dataset	r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation (sft_balanced)
Train Samples	10,000 (subset of 52K balanced SFT corpus)
Method	QLoRA (r=64, alpha=32, all linear layers)
Effective Batch Size	16
Learning Rate	2e-4 (cosine schedule)
Max Context	4,096 tokens
Final Loss	0.59–0.67
Hardware	Kaggle T4 x 2

Limitations

Trained on 10K samples; may underperform full 52K variant on edge cases
Reasoning quality degrades beyond 4K context window
Inherits base MiniCPM5 knowledge cutoff (~2024)
May produce verbose sections when concise answers are preferred

License

Apache-2.0 (inherited from base MiniCPM5-2B). Verify distillation dataset license terms before commercial use.

Acknowledgments

OpenBMB for MiniCPM5-2B
r0b0tlab for the multi-teacher distillation dataset
Unsloth for accelerated training infrastructure

Downloads last month: -

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for jigs97022/minicpm5-2b-distilled-reasoning

Base model

openbmb/MiniCPM5-2B

Finetuned

(15)

this model

Quantizations

2 models