MiniCPM5-1B-AX637

MiniCPM5-1B on AXERA AX637

Ready-to-run, text-only deployment package for openbmb/MiniCPM5-1B on an AX637 aarch64 board.

• Runtime: packaged axllm binary with OpenAI-compatible HTTP API and CLI.
• Target: AX637.
• Context profile: kv_cache_len=1024, prefill_len=128, and maximum prefill capacity 896 tokens.
• Included assets: 24 decoder .axmodel files, post-processing .axmodel, embedding weights, tokenizer, runtime configuration, and bin/axllm.

Supported Platform

• AX637 aarch64 board with the AXERA runtime installed.

This is a text-only package. The packaged configuration has enable_thinking=false.

Performance

Measurements below were taken on an AX637 board with the packaged runtime. TTFT means time to first generated token.

The long validation request generated five tokens and exercised every shipped prefill history group: 0, 128, 256, 384, 512, 640, and 768. Actual latency depends on board memory pressure, prompt length, and output length.

Startup Runtime Footprint

The CMM figure is the AX Engine CMM-pool delta measured from before startup to after all 24 decoder layers and the post model were loaded. The OS-memory figure is the board-side process resident set size (VmRSS) after the same startup point. VmSize is shown separately because the package uses memory mapping; it is virtual address space, not physical OS memory. CMM and RSS are the startup consumption values to use when budgeting a board, while actual system availability depends on other workloads.

Context Layout

Prompts longer than 128 tokens are split into chunks. The runner selects the smallest compatible prefill group for each chunk. Leave room inside the 1024-token KV window for generated tokens when sending long prompts.

Package Layout

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

This is a flat runtime package. Run axllm from the package root; it reads the root-level tokenizer, configuration, embedding, and .axmodel files directly.

Download the Package

Download this repository on the host that will transfer or mount it on the board:

mkdir -p AXERA-TECH/MiniCPM5-1B-AX637
cd AXERA-TECH/MiniCPM5-1B-AX637
hf download AXERA-TECH/MiniCPM5-1B-AX637 --local-dir .

Run on the Board

The package includes a validated AX637 axllm binary. From the package root:

chmod +x ./bin/axllm
export LD_LIBRARY_PATH=/opt/lib:${LD_LIBRARY_PATH:-}
./bin/axllm serve . --port 8000

The service exposes:

GET  http://<board-ip>:8000/health
GET  http://<board-ip>:8000/v1/models
POST http://<board-ip>:8000/v1/chat/completions

Expected model identifier:

AXERA-TECH/MiniCPM5-1B-AX637-C128-P896-CTX1024

Verify readiness:

curl http://127.0.0.1:8000/health
curl http://127.0.0.1:8000/v1/models

Example health response:

{
  "concurrency": 0,
  "max_concurrency": 1,
  "status": "healthy"
}

Text Request

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "AXERA-TECH/MiniCPM5-1B-AX637-C128-P896-CTX1024",
    "messages": [
      {
        "role": "user",
        "content": "中国的首都是哪里？请只回答城市名。"
      }
    ],
    "max_tokens": 32,
    "temperature": 0
  }'

The response uses the standard OpenAI chat-completions JSON shape. Set the OpenAI client base URL to http://<board-ip>:8000/v1 and use the model identifier shown above.

Interactive CLI

export LD_LIBRARY_PATH=/opt/lib:${LD_LIBRARY_PATH:-}
./bin/axllm run .

Type /q or /exit to leave the interactive session.

Conversion References

If you need the original model files or want to rebuild the deployment artifacts, start with:

• Original Hugging Face model: openbmb/MiniCPM5-1B
• AXERA conversion and deployment workflow: AXERA-TECH/MiniCPM5-1B.axera

Discussion

• GitHub Issues
• QQ group: 139953715