Qwen3 235B A22B is a massive mixture-of-experts model that activates only 22 billion parameters at a time, keeping inference costs manageable despite its enormous total capacity. It supports a hybrid thinking mode — toggling between deliberate step-by-step reasoning and faster direct responses depending on the task. The model handles multilingual workloads well and tends to be thorough, sometimes verbose, when working through complex problems.
| Benchmark | Score | Type | Recorded |
|---|---|---|---|
| LiveCodeBench | 70.7 | accuracy | 5d ago |
| AIME 2024 | 85.7 | accuracy | 5d ago |
| AIME 2025 | 81.5 | accuracy | 5d ago |
| Aider Polyglot | 59.6 | accuracy | 5d ago |