openbmb/ minicpm5-2b:q6_K

1,469 Downloads Updated 2 days ago

SOTA on-device LLMs, small yet powerful

thinking 2b

ollama run openbmb/minicpm5-2b:q6_K

curl http://localhost:11434/api/chat \
  -d '{
    "model": "openbmb/minicpm5-2b:q6_K",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='openbmb/minicpm5-2b:q6_K',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'openbmb/minicpm5-2b:q6_K',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 2 days ago

2 days ago

78abc56d4a29 · 2.1GB ·

model

archllama

·

parameters2.52B

·

quantizationQ6_K

2.1GB

template

<s>{{- if .Messages }}{{- range $i, $_ := .Messages }}{{- $last := eq (len (slice $.Messages $i)) 1

692B

system

You are a helpful assistant.

28B

params

{ "num_ctx": 4096, "stop": [ "[\"<|im_start|>\",\"<|im_end|>\",\"</s>\"]" ],

132B

Readme

SOTA on-device LLMs, small yet powerful

MiniCPM5-2B 🤗 🤖 | MiniCPM5-1B | Cookbook | GitHub

MiniCPM5-2B is the second model in the MiniCPM5 series. It is a dense 2.5B LlamaForCausalLM built for on-device assistants, coding agents, tool use, and reasoning. Official Ollama can run it directly — no custom fork.

🏆 2B-class open-source SOTA. Average 53.9 in our comparison set, ahead of strong 2B models (LFM2.5-2.6B, Qwen3.5-2B, Gemma-4-E2B-it) and also above the 4B-class models listed there (best 4B: 51.1). Gains are clearest in code, math, long context, tool use, and agent tasks.
🧠 Hybrid reasoning. The same checkpoint supports Think / No-Think. Use Ollama’s think switch, or /set think in the CLI.
🛠️ Tools and long context. Native tool calling, 128K context. Default num_ctx is 4096; raise it for long documents.
📂 Open data. Training data is released in the UltraData family (UltraX, UltraData-Code, UltraData-SFT-Agent-2609, UltraData-RL-2609).
💫 Easy on-device use. Default tag is Q4_K_M (latest / 2b, 1.6GB). Also available: q4_0, q4_1, q4_K_S, q5_0, q5_1, q5_K_M, q5_K_S, q6_K, q8_0, f16.

Quick start

”`bash ollama run openbmb/minicpm5-2b ollama run openbmb/minicpm5-2b –think ollama run openbmb/minicpm5-2b:q8_0

# SOTA on-device LLMs, small yet powerful

[MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B) 🤗 🤖 | [MiniCPM5-1B](https://ollama.com/openbmb/minicpm5) | [Cookbook](https://github.com/OpenSQZ/MiniCPM-V-CookBook) | [GitHub](https://github.com/OpenBMB/MiniCPM)

**MiniCPM5-2B** is the second model in the MiniCPM5 series. It is a dense 2.5B `LlamaForCausalLM` built for on-device assistants, coding agents, tool use, and reasoning. Official Ollama can run it directly — no custom fork.

- 🏆 **2B-class open-source SOTA.** Average **53.9** in our comparison set, ahead of strong 2B models (LFM2.5-2.6B, Qwen3.5-2B, Gemma-4-E2B-it) and also above the 4B-class models listed there (best 4B: 51.1). Gains are clearest in code, math, long context, tool use, and agent tasks.
- 🧠 **Hybrid reasoning.** The same checkpoint supports Think / No-Think. Use Ollama's think switch, or `/set think` in the CLI.
- 🛠️ **Tools and long context.** Native tool calling, 128K context. Default `num_ctx` is 4096; raise it for long documents.
- 📂 **Open data.** Training data is released in the [UltraData](https://ultradata.openbmb.cn/) family (UltraX, UltraData-Code, UltraData-SFT-Agent-2609, UltraData-RL-2609).
- 💫 **Easy on-device use.** Default tag is **Q4_K_M** (`latest` / `2b`, 1.6GB). Also available: `q4_0`, `q4_1`, `q4_K_S`, `q5_0`, `q5_1`, `q5_K_M`, `q5_K_S`, `q6_K`, `q8_0`, `f16`.

## Quick start

```bash
ollama run openbmb/minicpm5-2b
ollama run openbmb/minicpm5-2b --think
ollama run openbmb/minicpm5-2b:q8_0

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