AlibabaQwen3 235B A22B 2507
Qwen3 235B A22B 2507 is an Alibaba open-weights reasoning model, released 22 Jul 2025. Its best configuration (thinking reasoning effort) ranks #247 of 372 on the BenchLeader Index at 52.6 ±3.7, in the lower half. It scores highest in maths (62) and lowest in composite (45). At $0.747 per million tokens blended it is mid-priced. Output speed of 61 tokens per second puts it slower than most, with a first answer in 35.7 s. It has been measured at 2 reasoning-effort settings; tables show the best-scoring one. Last measured 2 Sept 2026.
- Blended price
- $0.747/M
- $0.230 in · $2.30 out
- Output speed
- 61 tok/s
- measured by Artificial Analysis
- First answer
- 36 s
- first token 0.57 s
- Context
- 256k
- Overall index53
- Reasoning59
- Coding50
- Agents & tools46
- Maths62
- Knowledge45
- Instruction following52
- Human preference59
- Long context60
- Composite45
Versions
Alibaba has shipped 2 models under this name. Each is ranked on its own results; a newer version often has fewer results so far, which holds its index nearer the average until more arrive.
| Model | Released | Index | Rank |
|---|---|---|---|
| Qwen3 235B A22B 2507thinkingthis page | 22 Jul 2025 | 52.6 | #247 |
| Qwen3 235B A22Bthinking | 28 Apr 2025 | 44.8 | #417 |
Reasoning-effort configurations
The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Default” means the publisher did not say which setting was used.
| Effort | Index | Rank | Speed | First answer | Chat reply cost | Categories |
|---|---|---|---|---|---|---|
| thinkingbest | 52.6 | #247 | 61 tok/s | 36 s | $0.0008 | Agents & tools 46 · Coding 50 · Composite 45 · Human preference 59 · Instruction following 52 · Knowledge 45 · Long context 60 · Maths 62 · Reasoning 59 |
| default | 51.0 | #281 | 58 tok/s | 2.38 s | $0.0004 | Agents & tools 57 · Coding 52 · Composite 44 · Human preference 61 · Instruction following 48 · Knowledge 43 · Long context 43 · Reasoning 62 |
Benchmark results
One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.
Reasoning
| Benchmark | thinking | default | Source |
|---|---|---|---|
| GPQA Diamond | 80.0%#115 | – | Epoch AI Benchmarking Hub |
| LMArena Hard Prompts | 1417#138 | 1448#94 | LMArena |
| GPQA Diamond (AA)not in index | 79.0%#200 | 75.3%#243 | Artificial Analysis |
| Humanity's Last Exam (AA)not in index | 15.9%#193 | 11.1%#240 | Artificial Analysis |
Coding
| Benchmark | thinking | default | Source |
|---|---|---|---|
| SciCode | 42.4%#99 | – | SciCode |
| WeirdML | 41.0%#98 | 38.7%#111 | WeirdML |
| LMArena Coding | 1443#135 | 1472#98 | LMArena |
| SciCode (AA)not in index | 41.4%#111 | – | Artificial Analysis |
Agents & tools
| Benchmark | thinking | default | Source |
|---|---|---|---|
| Terminal-Bench Hard | 13.6%#195 | 15.2%#187 | Artificial Analysis |
| τ²-Bench Telecom (AA)not in index | 53.2%#186 | 33.3%#238 | Artificial Analysis |
| BFCL Overall | – | 52.1%#21 | Berkeley Function Calling Leaderboard |
Maths
| Benchmark | thinking | default | Source |
|---|---|---|---|
| OTIS Mock AIME | 86.7%#73 | – | Epoch AI Benchmarking Hub |
Knowledge
| Benchmark | thinking | default | Source |
|---|---|---|---|
| SimpleQA Verified | 40.4%#40 | – | Epoch AI Benchmarking Hub |
| AA-Omniscience | -46.5#297 | -44.0#284 | Artificial Analysis |
| MultiNRC | 27.1%#28 | – | Scale AI SEAL |
Instruction following
| Benchmark | thinking | default | Source |
|---|---|---|---|
| IFBench | 51.2%#166 | 46.0%#195 | Artificial Analysis |
Human preference
| Benchmark | thinking | default | Source |
|---|---|---|---|
| LMArena Text | 1400#133 | 1423#103 | LMArena |
Long context
| Benchmark | thinking | default | Source |
|---|---|---|---|
| Fiction.LiveBench 120k | 68.8%#8 | – | Fiction.live |
| AA-LCR | 72.0%#144 | 33.9%#312 | Artificial Analysis |
Composite
| Benchmark | thinking | default | Source |
|---|---|---|---|
| AA Intelligence Index | 12.7#264 | 12#276 | Artificial Analysis |
Where to run it
Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.
| Provider | Speed | First token | Input $/M | Output $/M | Context | Quantisation |
|---|---|---|---|---|---|---|
| Google Vertex (US) (ZDR) | 40 tok/s | 0.58 s | $0.220 | $0.880 | 262k | – |
| Alibaba Cloud Int. | 39 tok/s | 0.51 s | $0.149 | $0.598 | 131k | – |
| Google Vertex (US) (ZDR) | 38 tok/s | 0.51 s | $0.250 | $1.00 | 262k | – |
| AtlasCloud | 31 tok/s | 1.01 s | $0.200 | $0.880 | 131k | fp8 |
| Nebius Token Factory | 31 tok/s | 0.70 s | $0.200 | $0.600 | 262k | fp8 |
| NovitaAI | 29 tok/s | 0.60 s | $0.090 | $0.580 | 131k | fp8 |
| StreamLake | 27 tok/s | 0.95 s | $0.210 | $0.840 | 128k | – |
| Parasail | 23 tok/s | 0.57 s | $0.140 | $0.800 | 131k | fp8 |
| Venice | 18 tok/s | 0.69 s | $0.150 | $0.750 | 128k | fp8 |
| DeepInfra | 12 tok/s | 0.32 s | $0.090 | $0.550 | 262k | fp8 |
| GMICloud | 10 tok/s | 1.30 s | $0.087 | $0.350 | 262k | fp8 |
Price history
Listed price per 1M tokens over time, as recorded by OpenRouter for the provider with the longest history.
What a task costs
Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache. Reasoning tokens are not modelled.
| Workload | Tokens in / out | Cost | With caching | Time |
|---|---|---|---|---|
| Chat reply | 400 / 300 | $0.0008 | – | 40.6 s |
| Summarise a 30-page report | 12,000 / 600 | $0.0041 | – | 45.5 s |
| Code edit | 6,000 / 1,500 | $0.0048 | – | 1.0 min |
| Agentic coding session | 60,000 / 4,000 | $0.023 | – | 1.7 min |
| Structured extraction | 2,000 / 200 | $0.0009 | – | 39.0 s |
See also
Data as of 10 Sept 2026. Compare these configurations.