AXERA-TECH / MiniCPM5-1B

MiniCPM5-1B on AXERA NPU

Ready-to-run AX650 deployment package for openbmb/MiniCPM5-1B .

• Runtime: axllm serve and axllm run
• Target: AX650 / AX650N aarch64
• Model type: text-only LLM
• Context layout: prefill_len=128 , kv_cache_len=2047 , prefill_max_token_num=1280
• Default generation mode: enable_thinking=true
• Included board binary: bin/axllm

This package has been validated through the OpenAI-compatible axllm serve API.

Supported Platform

• AX650 / AX650N

Capability

MiniCPM5-1B in this package is a pure text model.

• Text chat: supported
• Multi-turn text chat: supported
• Image, video, and audio inputs: not supported

Performance and Footprint

Measurements below were taken on AX650 on 2026-06-06 with the packaged axllm binary. TTFT means time to first token.

The text smoke prompts were kept within one 128 -token prefill group.

Flash Footprint

Startup Runtime Footprint

The startup log records remain_cmm(6692 MB) after the post axmodel was initialized.

Runtime layout from the same log:

Package Layout

.
├── README.md
├── bin/
│   ├── axllm
│   └── axllm.version.json
├── llama_p128_l0_together.axmodel
├── ...
├── llama_p128_l23_together.axmodel
├── llama_post.axmodel
├── model.embed_tokens.weight.bfloat16.bin
├── config.json
├── post_config.json
└── minicpm5_tokenizer.txt

The runtime files are placed at the repository root so the package can be served directly by axllm .

Run on AX650

From the repository root on the board:

chmod +x ./bin/axllm
./bin/axllm serve .

Expected model id:

AXERA-TECH/MiniCPM5-1B-AX650-C128-P1152-CTX2047

Health check:

curl http://127.0.0.1:8000/health

Chat Request

Default behavior uses thinking mode because config.json sets enable_thinking=true .

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "AXERA-TECH/MiniCPM5-1B-AX650-C128-P1152-CTX2047",
    "messages": [
      {"role": "user", "content": "1+1等于几？请简短回答。"}
    ],
    "max_tokens": 128,
    "temperature": 0
  }'

In thinking mode, the API returns explicit <think>...</think> markup so clients can render reasoning and final answer separately.

Disable Thinking Per Request

Use top-level enable_thinking=false :

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "AXERA-TECH/MiniCPM5-1B-AX650-C128-P1152-CTX2047",
    "messages": [
      {"role": "user", "content": "中国的首都是哪里？请只回答城市名。"}
    ],
    "enable_thinking": false,
    "max_tokens": 32,
    "temperature": 0
  }'

The compatible HuggingFace-style form is also accepted:

{
  "chat_template_kwargs": {
    "enable_thinking": false
  }
}

Notes

• Previous assistant <think>...</think> content is stripped from normal chat history before the next user turn, matching the official MiniCPM5 template behavior.
• If a thinking response reaches max_tokens before the model emits </think> , axllm closes the client-visible thinking block before finishing the response stream.
• This release does not include the original HuggingFace safetensors weights.