Nebius Token Factory
Mini toolsStep 3.5 Flash
By StepFun
Model family: Step
Step 3.5 Flash is introduced as StepFun’s most capable open source foundation model for agents, using a sparse Mixture of Experts so only a small subset of its 196B parameters is active per token. It employs 3 way multi token prediction to reach around 100 to 300 tokens per second and supports a 256K context window via sliding window attention. The model targets deep reasoning, coding and browser based agency, achieving strong scores on SWE Bench Verified, Terminal Bench 2.0 and several agentic benchmarks, while remaining deployable on high end consumer hardware under an Apache 2.0 license.
New
Text
Gen 7
Overview
Step 3.5 Flash is StepFun’s open source sparse MoE LLM, with 196B total parameters but 11B active per token, tuned for fast agentic reasoning, coding and long context work while staying efficient.
About StepFun
StepFun AI is a creative AI assistant platform that offers chat, search, coding help, and multimodal generation (images, audio, video), plus research models like Step3 and NextStep-1.
Website:
stepfun.ai
Tools using Step 3.5 Flash
No tools found for this model yet.
Last updated: February 4, 2026
