Qwen3.5 397B A17B
Qwen · released Feb 16, 2026
The Qwen3.5 series 397B-A17B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. It delivers...
Specification
- Context window
- 262K
- Max output
- 236K
- Knowledge cutoff
- Not stated
- Parameters
- 403.4B
- Licence
- apache-2.0
- Serving providers
- 10
- Moderated
- No
- Uptime
- 100.0%
Intelligence
18.4
63th percentile
Coding
48.2
Coding Index
Agentic
8.3
Agentic Index
Output speed
84 t/s
Median across providers
Latency
390ms
Time to first token
Cost per task
$0.31
Estimated
Benchmarks
Where the score comes from
The Intelligence Index is a composite. These are the underlying evaluations this model was actually measured on.
Evaluation scores
Percentage correct · higher is better
- τ²-bench (Telecom)95.6%
- GPQA Diamond89.3%
- IFBench78.8%
- AA-LCR (long context)77.3%
- SciCode44.8%
- Terminal-Bench Hard40.9%
- Humanity's Last Exam29.0%
An evaluation missing from this list was not run for this model — it is not a zero.
View as table
| Evaluation | Score |
|---|---|
| τ²-bench (Telecom) | 95.6% |
| GPQA Diamond | 89.3% |
| IFBench | 78.8% |
| AA-LCR (long context) | 77.3% |
| SciCode | 44.8% |
| Terminal-Bench Hard | 40.9% |
| Humanity's Last Exam | 29.0% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- Step 3.7 FlashStepFun19.5
- Qwen3.5-35B-A3BQwen19.3
- Claude Sonnet 4.5Anthropic19.3
- LongCat 2.0Meituan19.1
- MiniMax M2MiniMax18.6
- Claude Opus 4.1Anthropic18.6
- Qwen3.5 397B A17BQwen18.4
- Qwen3.6 35B A3BQwen18.2
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Step 3.7 Flash | 19.5 |
| Qwen3.5-35B-A3B | 19.3 |
| Claude Sonnet 4.5 | 19.3 |
| LongCat 2.0 | 19.1 |
| MiniMax M2 | 18.6 |
| Claude Opus 4.1 | 18.6 |
| Qwen3.5 397B A17B | 18.4 |
| Qwen3.6 35B A3B | 18.2 |
Percentile among all indexed models
Pricing
What it costs to run
List prices per million tokens, plus what one representative task works out to.
List price
- Input / 1M tokens
- $0.5
- Output / 1M tokens
- $3.6
- Cached input / 1M
- $0.3
- Blended 3:1
- $1.27
One task, estimated
$0.31
- Input tokens
- 50,000
- Output tokens
- 80,000
- Profile
- Reasoning
Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).
Serving providers
Speed and latency figures are medians across these providers, so a widely-served model reports a blend rather than any single endpoint.
Arena
Head-to-head generation quality
Elo from pairwise judgements, broken out by the kind of thing the model was asked to build.
- Overall Elo
- 1,192
- Win rate
- 52.5%
- Strongest at
- Websites
- Tournaments
- 5,253
Elo by category
Dot position on a 1,125–1,200 scale · Elo has no meaningful zero
- 1,192
- 1,176
- 1,165
- 1,161
- 1,143
- 1,137
Agent categories (full-stack apps, mobile apps) are only scored for models tested in agent harnesses.
View as table
| Category | Elo | Win rate |
|---|---|---|
| Websites | 1192 | 52.5% |
| Data visualisation | 1176 | 53.2% |
| UI components | 1165 | 51.4% |
| 3D scenes | 1161 | 56.7% |
| Game development | 1143 | 50.2% |
| SVG | 1137 | 56.1% |