Step 3.5 Flash: StepFun's 196B Model With 11B Active Parameters

What Step 3.5 Flash is

Step 3.5 Flash is StepFun's open-weight mixture-of-experts model: 196 billion total parameters with just 11 billion activated per token, and a 256K token context window — recorded here as 262.1K, the same number written out. It is roughly $0.10 per million input tokens and $0.30 output through third-party providers. On LMArena's text leaderboard Step 3.5 Flash has 1393 Elo (rank 174); by category, 1449 Elo in Coding (rank 160), 1407 Elo in Math (rank 155), 1429 Elo in Expert (rank 150) and 1345 Elo in Creative Writing (rank 186).

StepFun’s decoding-cost estimates

StepFun publish estimated decoding cost at a 128K context on Hopper GPUs, normalised against themselves. Step 3.5 Flash is the 1.0× baseline at 100 tokens per second with MTP-3 and EP8. Against it: MiMo-V2 Flash at 1.2×, MiniMax M2.1 at 3.9×, DeepSeek V3.2 at 6.0× (and only 33 tokens per second), and both Kimi K2 Thinking and GLM-4.7 at 18.9×. These are StepFun's own estimates under StepFun's chosen configuration.

StepFun’s benchmark figures

From StepFun's published comparison: on mathematics, AIME 2025 at 97.3 against DeepSeek V3.2's 93.1 and GLM-4.7's 95.7; HMMT February 2025 at 98.4 against 92.5 and 97.1; HMMT November at 94.0; IMOAnswerBench at 85.4. On coding, LiveCodeBench-V6 reaches 86.4, SWE-bench Verified 74.4 and Terminal-Bench 2.0 51.0. On agentic and search evaluations it records τ²-Bench 88.2, GAIA without files 84.5, xbench-DeepSearch 83.7 and ResearchRubrics 65.3. BrowseComp comes in at 51.6, rising to 69.0 with their Context Manager enabled, which describes a system rather than a bare model. All vendor-reported.

Architecture

StepFun's card lists a 45-layer Transformer with a hidden dimension of 4,096 and a vocabulary of 128,896 tokens, totalling 196.81 billion parameters (a 196B backbone plus a 0.81B head) with about 11 billion active per generated token. Each MoE layer has 288 routed experts plus one always-active shared expert, with the top 8 selected per token. The 256K context uses a 3:1 ratio of sliding-window attention to full attention: three sliding-window layers for each full-attention layer. A multi-token-prediction head, built from sliding-window attention and a dense feed-forward network, predicts 4 tokens in a single forward pass.

Speed, sampling and deployment

StepFun state a generation throughput of 100 to 300 tokens per second in typical use with three-way multi-token prediction, peaking at 350 tokens per second for single-stream coding. They recommend temperature 0.6 and top_p 0.95 for general chat, and temperature 1.0 and top_p 0.95 for reasoning and agent use. The model is released under Apache 2.0 and is available through StepFun's platform and OpenRouter, which StepFun note offered a free trial. For local inference StepFun list vLLM, SGLang, Hugging Face Transformers and llama.cpp, and name the Mac Studio M4 Max and NVIDIA DGX Spark as examples of supported hardware. A technical report is on arXiv (2602.10604).

Chinese browsing results

StepFun also report BrowseComp-ZH 66.9, or 73.7 with the Context Manager, and xbench-DeepSearch (October 2025) 56.3. With the Context Manager, the agent resets its context and restarts the loop once the effective context passes a threshold.

Official Resources