Step 3.7 Flash
StepFun · released May 28, 2026
Step 3.7 Flash is StepFun's latest high-efficiency multimodal Mixture-of-Experts model. It pairs a 196B-parameter language backbone with a vision encoder for native image and video understanding, activating roughly 11B parameters...
Specification
- Context window
- 262K
- Max output
- 256K
- Knowledge cutoff
- Not stated
- Parameters
- 201.4B
- Licence
- apache-2.0
- Serving providers
- 3
- Moderated
- No
- Uptime
- 100.0%
Intelligence
30.9
75th percentile
Coding
39.6
Coding Index
Agentic
21.7
Agentic Index
Output speed
153 t/s
Median across providers
Latency
253ms
Time to first token
Cost per task
$0.10
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)98.5%
- GPQA Diamond80.9%
- AA-LCR (long context)69.7%
- IFBench67.3%
- SciCode40.0%
- Terminal-Bench Hard35.6%
- Humanity's Last Exam21.4%
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) | 98.5% |
| GPQA Diamond | 80.9% |
| AA-LCR (long context) | 69.7% |
| IFBench | 67.3% |
| SciCode | 40.0% |
| Terminal-Bench Hard | 35.6% |
| Humanity's Last Exam | 21.4% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- Qwen3.6 35B A3BQwen32.1
- GPT-5.1-Codex-MiniOpenAI31.3
- Ring-2.6-1TInclusionAI31.1
- o3OpenAI31.1
- Step 3.7 FlashStepFun30.9
- Mistral Medium 3.5Mistral AI30.4
- Qwen3.5-35B-A3BQwen29.9
- Claude Sonnet 4.5Anthropic29.9
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Qwen3.6 35B A3B | 32.1 |
| GPT-5.1-Codex-Mini | 31.3 |
| Ring-2.6-1T | 31.1 |
| o3 | 31.1 |
| Step 3.7 Flash | 30.9 |
| Mistral Medium 3.5 | 30.4 |
| Qwen3.5-35B-A3B | 29.9 |
| Claude Sonnet 4.5 | 29.9 |
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
- $0.2
- Output / 1M tokens
- $1.15
- Cached input / 1M
- $0.04
- Blended 3:1
- $0.438
One task, estimated
$0.10
- 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).
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
- 1,211
- Win rate
- 45.8%
- Strongest at
- Websites
- Tournaments
- 10,795
Elo by category
Dot position on a 1,100–1,225 scale · Elo has no meaningful zero
- 1,211
- 1,204
- 1,198
- 1,191
- 1,177
- 1,115
Agent categories (full-stack apps, mobile apps) are only scored for models tested in agent harnesses.
View as table
| Category | Elo | Win rate |
|---|---|---|
| Websites | 1211 | 45.8% |
| UI components | 1204 | 43.7% |
| Data visualisation | 1198 | 45.3% |
| Game development | 1191 | 41.7% |
| 3D scenes | 1177 | 41.8% |
| SVG | 1115 | 38.6% |