Free Sorting of OpenClaw Repository Using Local AI Models
Hugging Face
Google/DeepMind
DeepSeek
Anthropic
The Hugging Face blog authors demonstrated how to use local Gemma and Qwen models in an agentic setup to sort issues and PRs in the OpenClaw repository in real time. The system uses a read-only Pi agent and notifies via Discord, which is cheaper than closed models like Claude or GPT-5. Tests on 330 examples showed Qwen is more accurate, while Gemma is faster, and both virtually eliminate false positives.
The authors of the Hugging Face blog described how they built a real-time task and PR sorting system for the OpenClaw repository using local open-weight models Gemma 4 26B and Qwen 3.6 35B within the Pi agent framework. The idea emerged after Anthropic removed its latest Claude Fable 5 model, confirming that closed models can disappear at any time. The system runs on a local NVIDIA GB10 with 128 GB of unified memory. For classification, an agent receives the title, body, and summary from a PR, and can only execute read-only commands via reposhell (a restricted bash) or call a final JSON with the result. This protects against prompt injections. The orchestration is built on gitcrawl for local mirroring and localpager for the job queue. Testing on 330 issues and PRs showed: Gemma achieves F1 0.800 with a time of 1.41 s/line and 402.6 output tokens/s, Qwen achieves F1 0.824 with 13.51 s/line and 145.3 tokens/s, and DeepSeek-V4-Flash (284B) achieves F1 0.811 but 144 s/line. False positives are 227 for Gemma, 106 for Qwen, and 30 for DeepSeek. The authors note that accuracy improves with NVFP4 quantization on Blackwell and vLLM optimizations, and the cost is limited to electricity.
Source: Hugging Face blog —
original