Step 3.5
#266 in Frontier-SpraakmodelleStepfun · v3.5 · siet 2026-02-02 · 5× · tolest 29. Juni 2026
2
Momentum
Step 3.5 Flash is an open-weight reasoning model by StepFun built on a sparse Mixture-of-Experts architecture, activating only 11B of its 196B total parameters per token. The model is purpose-built for coding and agentic tasks and supports a 256K-token context window. It is a text-only model without image or video modality and is released under the Apache 2.0 license. The API is accessible via the StepFun platform as well as third-party providers such as OpenRouter.
Momentum-Verloop
04.07.02.10.
Features
| Key-Benchmark (%) | IMO-AnswerBench: 85,4% / LiveCodeBench-v6: 86,4% / τ²-Bench: 88,2% / BrowseComp (mit Context Manager): 69,0% |
| Coding Benchmark (SWE-bench Pro %) | 74.4% on SWE-bench Verified; 51.3% on SWE-bench Pro (per comparison data for successor version Step 3.7 Flash) |
| Context Window (Tokens) | 256,000 tokens (256K) per official documentation; 262,144 tokens (262K) per OpenRouter API listing |
| Lizenz | Apache 2.0 (Open Weight, kommerzielle Nutzung erlaubt) |
| Multimodal (Image/Video) | No – text-only; no image or video input supported |
| Multimodalität | Nein – reines Textmodell, kein Bild-/Video-Input (Vision erst mit Step 3.7 Flash eingeführt) |
| NLP Benchmark Score | AIME 2025: 97.3%; LiveCodeBench v6: 86.4%; τ²-Bench: 88.2%; IMO-AnswerBench: 85.4% |
| Plattform | StepFun API (International & China), OpenRouter, Hugging Face, GitHub, DeepInfra, SiliconFlow, NVIDIA NIM, vLLM |
| Price per 1M Input Tokens | $0.09 (via OpenRouter) or $0.10 (via StepFun directly) |
| Preis pro 1M Token | 0,10 $ Input / 0,30 $ Output (StepFun/OpenRouter-Listenpreis) |
| Price Tier (Monthly) | No monthly subscription model; pure pay-per-use API. OpenRouter also offers a free trial access. |
| Release-Datum | 29. Januar 2026 |