Qwen 3 235B-A22B
Language Retired Full comparison ↗ alibaba/qwen-3-235b-a22b · by Alibaba
· mixture-of-experts
Pricing — 1 offering(s)
About this offering: Retired. OpenRouter stopped serving Qwen3 235B A22B on 2026-10-09 and recommends Qwen3.7 Plus instead. Its newer 2507 variants are separate models.
Input tokens
Retired — no longer sold · last price $0.455 / 1M tokens (input), until 2026-10-08
- $0.455 / 1M tokens (input) Historical 2025-04-28 → 2026-10-08
Output tokens
Retired — no longer sold · last price $1.82 / 1M tokens (output), until 2026-10-08
- $1.82 / 1M tokens (output) Historical 2025-04-28 → 2026-10-08
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Qwen 3 235B-A22B fits into a cost-aware routing setup
See how →Capability profile
How Qwen 3 235B-A22B rates across core capability dimensions, with the task-level evidence behind each rating.
Best-in-class among open-weights MoE models; thinking mode rivals frontier closed models.
SWE-bench Pro (1,865 multi-language GitHub issues across 41 repos — a harder, less contamination-prone successor to SWE-bench Verified; pass@1, no retries): resolves 21.4% — fails roughly 4 issues in 5. Independent (Scale AI public leaderboard, #18; SWE-agent scaffold). No SWE-bench Verified score has been published for this model.
Flexible — both structured non-thinking and extended-thinking modes available.
128K context.
Best open-weights multilingual coverage in this tier; excels in Chinese and Asian languages.
MoE efficiency but 22B active parameters is not fast-tier; comparable to dense 30B.
Coding benchmarks
View the full coding benchmark leaderboard →| Benchmark | Score | Harness | Source |
|---|---|---|---|
| livecodebench ↗ | 70.7% 2025-05 | — | 📄 paper ↗ |
| aider-polyglot ↗ | 59.6% 2025-04 | aider | 📊 leaderboard ↗ |
| swe-bench-pro | 21.4% 2026-01 | swe-agent | 📊 leaderboard ↗ |
Operator guidance
Exceptional open-weights value for reasoning workloads. At $0.455/$1.82 via OpenRouter, it competes with o4-mini on reasoning at ~40% of the cost. Use thinking mode for hard problems, non-thinking for high-volume workloads. The Qwen 3 235B-A22B-2507 variant is available at $0.09/$0.10 for lower cost.
Use cases
- Frontier-quality reasoning tasks at open-weights cost via hosted APIs
- Multilingual enterprise workloads needing a non-Chinese-lab option
- Self-hosted inference where reasoning capability is required
Limitations
- Context window (128K) smaller than Maverick or Scout
- Chinese lab; data-residency or compliance constraints may apply
- Capability ratings are qualitative, not a single cited benchmark run