MiniCPM-o 4.5 Deploy
v1.0.0Deploy MiniCPM-o 4.5 multimodal model via Web Demo, vLLM Serve, or llamacpp-omni. Use when the user asks to deploy, start, configure, or troubleshoot MiniCPM...
⭐ 0· 130·0 current·0 all-time
byDennis Huang@zmxjj
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
The name/description (deploy MiniCPM-o 4.5 via Web Demo / vLLM / llamacpp-omni) matches the included files and runtime instructions. The provided download/verification script, references, and step-by-step install/launch commands align with model download and deployment tasks. No unrelated credentials, binaries, or config paths are requested.
Instruction Scope
SKILL.md gives explicit installation and deployment steps (conda or venv, pip installs, git clone, running install.sh, starting gateway/workers). It instructs network downloads from GitHub, HuggingFace, and ModelScope and running a local Python script to verify/download models. This scope is appropriate for a deployment guide, but the agent will run networked downloads and install packages at runtime — actions that have real consequences and should be done in a trusted/isolated environment.
Install Mechanism
There is no packaged install spec (instruction-only) but the included scripts call pip to install dependencies on demand and recommend running the project's install.sh. That is expected for this type of skill but is higher-risk than pure documentation because it causes code to be written/executed locally. A non-official mirror (https://hf-mirror.com) is suggested in troubleshooting; using unofficial mirrors carries extra risk and should be avoided unless you trust the mirror.
Credentials
The skill declares no required environment variables or credentials. The runtime instructions mention optional environment variables (e.g., HF_ENDPOINT) and may prompt for HuggingFace/ModelScope credentials if private models are used, which is proportional. There are no requests for unrelated secrets or broad system credentials.
Persistence & Privilege
always is false and the skill does not request persistent platform-level privileges or modify other skills. It runs as-needed commands and scripts in the user's environment; autonomous invocation by the agent is allowed by default but not excessive here.
Assessment
This skill appears internally consistent for deploying MiniCPM-o 4.5, but it performs network downloads and installs packages and runs project install scripts. Before using it: (1) verify the upstream repository (the instructions clone https://github.com/OpenBMB/MiniCPM-o-Demo); (2) avoid using untrusted mirrors (the SKILL.md suggests hf-mirror.com, which is not an official HuggingFace host); (3) run install steps in an isolated environment (container or VM) or a dedicated virtualenv/conda env; (4) do not provide private tokens unless you understand why they’re needed (private model downloads may prompt for credentials); and (5) inspect install.sh and any downloaded scripts yourself if you need stronger assurance.Like a lobster shell, security has layers — review code before you run it.
latestvk97febme7ncs54r1fddmgs8pfn833vpr
License
MIT-0
Free to use, modify, and redistribute. No attribution required.