Qwen3 235B MoE
Alibaba·
LLMsopen-weightMulti-GPU Self-Host
Overview
Hybrid thinking model supporting 119 languages. Seamlessly switches between fast responses and deep chain-of-thought reasoning.
Capabilities and innovations
235B total / 22B active parameters (MoE)Hybrid thinking (fast + deep reasoning)119 languages supportedMCP and tool-use supportThinking mode toggle within single modelMassive multilingual coverage (119 langs)Thinking budget control4-stage post-training pipeline
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 70.2% | unsourced |
| Humanity’s Last Exam (no tools) | 17.3% | unsourced |
| SWE-bench Pro | 32.2% | unsourced |
| MMLU | 88.5% | unsourced |
Architecture and hardware
- Parameters
- 235B total, 22B active per token (MoE)
- Estimated VRAM at Q4
- ~136 GB, Multi-GPU Self-Host class
- Quantization formats
- GGUF, GPTQ, AWQ
- Recommended runtime
- Ollama
- License
- Apache 2.0
Links
More from Alibaba
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.