This workflow reorganizes a FreeDiDi deployment post. Speed and model-comparison claims are observations from that source and were not independently reproduced by LingjingHub.
Goal
Run a Holo 3.1 model through a local OpenAI-compatible llama.cpp endpoint, connect that endpoint to OpenClaw, and use the resulting agent for browser and desktop tasks without sending every prompt to a hosted model provider.
1. Select the right model size
The source links Holo 3.1 variants around 0.8B, 4B, 9B, and 35B-A3B. Choose according to actual GPU memory and workload. Vision-capable packages may require both the main GGUF model and an mmproj file. File naming varies by repository, so confirm the exact pair in the current model card rather than relying on a copied batch script.
2. Serve the model with llama.cpp
Download llama.cpp from its official repository and use a build that matches your operating system and GPU runtime. Start llama-server on 127.0.0.1 with a conservative context length, then confirm the OpenAI-compatible endpoint responds before connecting an agent. Keep the service local unless authentication and network controls have been configured.
3. Connect OpenClaw
Install OpenClaw only from its official documentation. In the provider configuration, point the base URL to the local llama.cpp endpoint, typically a localhost address ending in /v1, and select the exact served model name. Avoid piping an unreviewed remote script directly into a privileged shell; inspect the official installation path and release artifacts first.
4. Add capabilities gradually
Start with read-only or low-impact tasks such as opening a test page, collecting public information, or organizing local sample files. Add browser automation skills only from sources you have reviewed. A skill expands what the agent can do, so it should be treated like executable software rather than a harmless prompt.
5. Define confirmation gates
Require explicit confirmation before downloads, form submission, sending messages, purchases, account changes, file deletion, or access to private browser sessions. Use a separate browser profile and a test workspace for early runs. Do not expose passwords, cookies, API keys, or personal files to an agent simply because its model runs locally.
6. Measure the real outcome
Evaluate completion rate, time to finish, retries, GPU memory, latency, and incorrect actions on your own task set. The source reports fast browser automation relative to another local model, but this remains a hardware- and task-specific observation.
Success criteria
The local endpoint responds reliably, OpenClaw uses the intended Holo model, test tasks stay within the approved permissions, and failures can be stopped without changing real accounts or data.
Recovery notes
If OpenClaw cannot reach the model, verify the localhost port, /v1 path, model identifier, and server logs. If vision calls fail, check that the matching mmproj file is loaded. Reduce context length or quantization size when memory is exhausted.
References
https://huggingface.co/collections/Hcompany/holo31
https://github.com/ggml-org/llama.cpp
Source article