Self-hosted GitLab review agent on 2x RTX 4090 running Qwen3.6-35B-A3B, 70% of comments resolved
Professional-Pear351 · reddit · 2026-10-21
A developer built a fully local GitLab code review agent on LangGraph.js, running Qwen3.6-35B-A3B (4-bit AWQ) via sglang on two RTX 4090s over PCIe with TP=2. Decode hits 180 tok/s single-stream at 50k context, prefill 9-10k tok/s, and 985 tok/s total at 8 concurrent requests. Each review averages 26 LLM calls and 2.3M tokens; after 1,300+ reviews, 70% of its comments get resolved by authors. It sometimes catches mistakes in code written by Claude Sonnet/Opus — mixing model families gives different perspectives. Main pain point: local models looping on themselves, ending 7% of runs despite guardrails. MIT-licensed and also works with Ollama, OpenRouter, Gemini and Groq.
More from coding & agent
- EAU: Co-evolving user simulator, environment, and RL setup for training LLM agents — UAlberta · 2026-10-21
- Meta's Self-Credit: Dense Rewards from Sparse RLVR via Intra-group Rescaling — meta · 2026-10-21
- Tamper-proof agent logs still can't prove the agent did the job: who reads the record back? — ZealousidealAct9661 · 2026-10-21
- One prompt: Claude builds an animated mech spider in Blender via MCP, first try — sidahuj · 2026-10-21
- OpenAI Details Coordinated Disclosure of Critical Vulnerabilities Found by Agents — deliciousmark12 · 2026-10-21
- Dev: coding agents make every idea 10 minutes away — and that's the problem — feregri_no · 2026-10-21