AiderFrom the builder
Qwen3 benchmark results
Aider compiled results for the Qwen3 models on its polyglot coding benchmark, across the diff and whole edit formats and several API providers, noting that providers serving open models can use very different inference settings and quantizations. With VLLM, bfloat16 and the recommended /no_think settings, Qwen3-235B-A22B scored 65.3% in whole format, and 61.8% through Alibaba's official API with thinking off. In diff runs on OpenRouter across all providers with default thinking settings, it scored 49.8% and Qwen3-32B scored 40.0%.
This article is being updated as new results become available.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.