treyleo16/ minicpm5

14 Downloads Updated 6 days ago

A 2B model with advanced reasoning. Made by OpenBMB on Hugging Face under Apache 2.0.

ollama run treyleo16/minicpm5

curl http://localhost:11434/api/chat \
  -d '{
    "model": "treyleo16/minicpm5",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='treyleo16/minicpm5',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'treyleo16/minicpm5',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size / Usage

Context

Input

minicpm5:latest

1.6GB · 128K context window · Text · 6 days ago

minicpm5:latest

1.6GB

128K

Text

Readme

MiniCPM5-2B

A dense 2B transformer from OpenBMB, built for on-device and local deployment. Repacked GGUF for Ollama.

ollama run treyleo16/minicpm5

Why this one

2B-class open-source SOTA. In OpenBMB’s own evals it averages 53.9, which beats every 4B-class model they benchmarked against (top score there was 51.1 — Qwen3.5-4B). Strongest in code reasoning, math, long-context, tool use, and agentic tasks.

Good fit for local coding agents, tool-calling workflows, and anything where you’d rather not hit a remote API.

Specs


Params	2,516,756,480 (1.98B non-embedding)
Layers	42
Attention	GQA — 16 Q heads, 2 KV heads
Context	131,072
Architecture	`LlamaForCausalLM`
Languages	English, Chinese
Quant
Size

Recommended params

PARAMETER temperature 0.7
PARAMETER top_p 0.95
PARAMETER stop "<|im_end|>"
PARAMETER stop "</s>"

Long context

Before ollama serve, if you’re pushing the context window:

export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0

Source

Weights: openbmb/MiniCPM5-2B-GGUF License: Apache-2.0 · Repacked by treyleo16

# MiniCPM5-2B

A dense 2B transformer from [OpenBMB](https://github.com/OpenBMB/MiniCPM),
built for on-device and local deployment. Repacked GGUF for Ollama.

```bash
ollama run treyleo16/minicpm5
```

## Why this one

2B-class open-source SOTA. In OpenBMB's own evals it averages **53.9**,
which beats every 4B-class model they benchmarked against (top score
there was 51.1 — Qwen3.5-4B). Strongest in code reasoning, math,
long-context, tool use, and agentic tasks.

Good fit for local coding agents, tool-calling workflows, and anything
where you'd rather not hit a remote API.

## Specs

| | |
|---|---|
| Params | 2,516,756,480 (1.98B non-embedding) |
| Layers | 42 |
| Attention | GQA — 16 Q heads, 2 KV heads |
| Context | 131,072 |
| Architecture | `LlamaForCausalLM` |
| Languages | English, Chinese |
| Quant |  |
| Size |  |

## Recommended params

```
PARAMETER temperature 0.7
PARAMETER top_p 0.95
PARAMETER stop "<|im_end|>"
PARAMETER stop "</s>"
```

## Long context

Before `ollama serve`, if you're pushing the context window:

```bash
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0
```

## Source

Weights: [openbmb/MiniCPM5-2B-GGUF](https://huggingface.co/openbmb/MiniCPM5-2B-GGUF)
License: Apache-2.0 · Repacked by [treyleo16](https://ollama.com/treyleo16)

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)