Qwen3-235B
AlibabaTEXT Open weight
- Provider
- Alibaba
- Modality
- TEXT
- Parameters
- 235B
- Context window
- 262,144 tokens
- Weights
- Open
- Released
- 22 Jul 2025
Qwen3-235B-A22B-Instruct is the mid-2025 open MoE instruct flagship (235B/22B, 256K) — a widely-deployed, self-hostable Apache-2.0 workhorse.
Best for
- Strong open non-reasoning instruct flagship
- Instruction-following, multilingual and long-context tasks
- Widely deployed and well-supported across the open ecosystem
- Pair with its Thinking sibling for hard reasoning
How it compares — and the India angle
- Apache 2.0 and only 22B active params — cost-effective to self-host on Indian cloud GPUs.
- Cheaper than US frontier APIs whether self-hosted or via OpenRouter/DashScope.
- Don't confuse with the Thinking-2507 variant, which scores much higher on math/GPQA.
Benchmarks
Representative public scores (approximate, higher is better) for relative comparison. Check the provider for the latest official results.
GPQA Diamond78
AIME (math)70
MMLU-Pro83
How to access
Free to self-host (Apache 2.0). Via OpenRouter/DashScope ~$0.20–0.70 / 1M input. 22B active / 235B total MoE.
Access Qwen3-235B