In July 2025, Alibaba launched Qwen3-235B-A22B, the largest MoE model in the Qwen3 family, supporting up to 256K context.
Key specs
- Total parameters: 235B, active parameters: 22B (MoE architecture)
- Context length: up to 256K tokens
- Target: long-document understanding, codebase analysis, multi-turn agent conversations
Why it matters
A 256K context window lets the model process roughly 20-30 thousand English tokens’ worth of material in one pass (estimate; check official docs for exact limits), which is practically meaningful for codebase-level understanding. The 22B active parameter count keeps inference cost well below a dense model of comparable size — the usual MoE advantage on the serving side.
What still needs verification
- Official benchmark comparisons against K2, DeepSeek, and Claude have been published, but the methodology and reproduction paths are not fully public; treat conclusions as pending independent evaluation.
- “256K” is the maximum window; usable length depends on GPU memory, the inference framework, and attention implementation. Load-test with real workloads before deployment.
