Qwen3-235B-A22B

Qwen3-235B-A22B is the flagship of Alibaba's Qwen3 family: a Mixture-of-Experts model with 235B total parameters but only 22B activated per token, released under Apache 2.0 (license verified on the HF repo). Its signature feature is hybrid thinking — a switchable reasoning mode with controllable thinking budgets (also via /think and /no_think prompt switches) — plus support for 119 languages and dialects, tool use/MCP, and strong agentic coding. The July 2025 "2507" refresh (Qwen3-235B-A22B-2507) substantially improved benchmark scores and context length. Deploy via vLLM, SGLang, Ollama, LM Studio, or llama.cpp.

Provider-reported (Qwen's official eval tables, 2507 releases): Thinking mode — AIME25 92.3, HMMT25 83.9, LiveCodeBench v6 74.1, CFEval 2134, GPQA 81.1, Arena-Hard v2 79.7; Instruct (non-thinking) mode — AIME25 70.3, LiveCodeBench v6 51.8, GPQA 77.5, ARC-AGI 41.8. Qwen claims the flagship is competitive with DeepSeek-R1, o3-mini and Gemini-2.5-Pro — treat head-to-head framing as provider-claimed. Independent: third-party reviews (Medium, dev.to) reproduced and corroborated the headline scores; LMArena coverage exists but I did not independently verify a current Elo figure.