MiniCPM5 Skills

MiniCPM5 Skills is the official SKILL.md pack inside OpenBMB/MiniCPM. The Hugging Face MiniCPM5-2B card points coding agents at these skills instead of improvising vLLM flags. On 2026-09-08 the skills/ tree listed 17 folders: a deploy router, backend cookbooks (transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, ArcLight, vLLM-Ascend), and finetune skills (TRL, LLaMA-Factory, ms-swift, Unsloth, xtuner, GGUF LoRA). The parent repo is Apache-2.0 and showed about 10,360 stars. This is a SKILL.md pack for Cursor / Claude Code style agents, not a YouMind marketplace Skill.

Compare MiniCPM5-2B if you wanted the weights, Claude Skills if you still needed how skills load, or Marketing Skills if you wanted a marketing library instead of a deploy router.

Key Features

• Router first: minicpm5-deploy asks for checkpoint, hardware, and goal, then invokes exactly one backend skill. It tells the agent not to inline commands.
• Backend matrix: CUDA server to vLLM, Ascend NPU to vLLM-Ascend, Apple Silicon to Ollama or MLX, CPU/low-VRAM to llama.cpp Q4KM.
• Finetune skills: TRL+PEFT, LLaMA-Factory, ms-swift, Unsloth, xtuner, GGUF LoRA. Each pairs with a one-page cookbook under docs/.
• Shared pitfalls in the router: MiniCPM5-2B think-mode sampling (temperature=1.0, top_p=0.95), 128k context flags, untied lm_head that breaks old MLX converters.
• Sanity check: after any backend, curl /v1/chat/completions with 1+1=? and expect HTTP 200.

Limitation: these skills only know MiniCPM5-1B and MiniCPM5-2B. They will not deploy Qwen3.8-27B or DeepSeek. Loading the whole pack without a checkpoint still wastes context.

Use Cases

• Claude Code / Cursor users who just pulled openbmb/MiniCPM5-2B and do not want to memorize engine flags.
• Apple Silicon laptops that should land on Ollama or MLX instead of a CUDA recipe.
• Fine-tune jobs that need LoRA on TRL or LLaMA-Factory without a generic "write a trainer" prompt.

Pricing

Getting Started

1. Open github.com/OpenBMB/MiniCPM/tree/main/skills.
2. Copy the skill folders into your agent's skill directory, or point the agent at skills/minicpm5-deploy/SKILL.md.
3. Give MODEL_PATH (openbmb/MiniCPM5-2B or a local dir with config.json).
4. Confirm MiniCPM5-2B if you still needed the card.

First-party resource: minicpm5-deploy SKILL.md.

Frequently Asked Questions

Same as the model page?

No. MiniCPM5-2B is the checkpoint. This pack is the agent instructions for serving and finetuning it.

Do I install with npx skills add?

The MiniCPM5-2B card links GitHub paths, not an npx skills add one-liner. Copy skills/ or clone the repo.

Will it pick vLLM on a Mac?

No. The matrix sends Apple Silicon to Ollama (easiest) or MLX (fastest).

Alternatives

• MiniCPM5-2B: the weights these skills assume.
• Claude Skills: how skills load in Claude Code.
• Marketing Skills: CRO and SEO, not model serving.

Tips

1. Quote 17 skill folders, Apache-2.0, and 10,360 repo stars from pages checked 2026-09-08.
2. Start at minicpm5-deploy, do not paste a vLLM command from memory.
3. Pass --max-model-len 131072 only if VRAM can hold 128k.

Conclusion

MiniCPM5 Skills is the official OpenBMB router so an agent deploys MiniCPM5 on the right engine instead of guessing flags. Start at the skills tree, then decide whether MiniCPM5-2B already covers the model you wanted.