Step 3.5 Flash: benchmark results
StepFun sparse MoE with approximately 11B active parameters and 256K context. The publisher specifies 196.81B architecture parameters (196B backbone plus 0.81B head); the configuration declares three next-token prediction layers, and the archived tensor total differs. Provider: StepFun. Released 2026-02-02. Access: Open.
Aggregate Rating 1670 ± 8, rank #270 of 1656 rated models, from 312 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenEvals - AIME 2026 | 96.67 | Accuracy (%) | 100 |
| WebTestBench - Checklist Coverage | 66 | Checklist coverage (%): share of gold test items the agent's | 100 |
| MathArena - BRUMO 2025 | 100 | Accuracy (%) | 98.9 |
| MathArena - HMMT Feb 2025 | 98.33 | Accuracy (%) | 98 |
| OpenCompass Research - LiveCodeBench v6 | 83.9 | Score (%) | 97.9 |
| MathArena - CMIMC 2025 | 95.62 | Accuracy (%) | 97.1 |
| OpenCompass Research - AIME 2025 | 95.7 | Score (%) | 95.7 |
| MathArena - AIME 2025 | 98.33 | Accuracy (%) | 95.1 |
| RewardBench 2 Math | 89.34 | Score (%) | 95 |
| ALL Bench LLM | 77.28 | Average Numeric Benchmark Score (%) | 94.7 |
| AA TAU-2 Bench | 94.44 | Accuracy (%) | 92.9 |
| Tau2-Bench Telecom | 94.44 | Success Rate (%) | 92.9 |
Interactive version: theaggregate.ai/model?slug=step-3-5-flash · How It Works · Data refreshed daily, snapshot 2026-10-20.