Developer Compares Local Qwen3.8-Flash-Next with Claude for Code Reviews

A developer successfully utilized a locally hosted Qwen3.8-Flash-Next model to conduct code reviews via Claude Code, comparing its performance directly against Claude Opus. The local model successfully caught issues that Claude missed, achieving accurate findings with zero false positives. This comparison highlights the growing viability of running powerful, large-scale open-weights AI models locally on high-end hardware for complex development tasks. It demonstrates that local setups can effectively rival cloud-based proprietary systems like Claude for intensive workflows such as code reviews. Running Qwen3.8-Flash-Next UD-Q4_K_XL on AMD hardware via llama-server yielded 14 to 21 tokens per second depending on context length, while consuming about 110K context tokens over 53 turns. Technical workarounds were required, including resolving MTP crash issues and overriding chat templates to handle mid-conversation system messages from Claude Code.