StepFun

Flagship agentic foundation model released September 20, 2026: sparse MoE with 600B total / 27B active parameters, 1M-token context, native text+image+video input and 64K max output at $1.00/$2.70 per 1M tokens; open weights announced for October 15, 2026.

Speech and audio model series launched September 15, 2026: Step-Audio 3 Realtime delivers full-duplex low-latency voice dialogue with adaptive reasoning, alongside ASR, TTS, Music and Gen variants that generate complete soundscapes from a single description.

Step Video TI2V is an video generation model from Stepfun, designed for video generation and motion content production.

Step Video T2V is an video generation model from Stepfun, designed for video generation and motion content production.

Step Video T2V Turbo is an video generation model from Stepfun, designed for video generation and motion content production.

Step 2 is an text generation model from Stepfun, designed for general AI application workflows.

Step 2 Mini is an text generation model from Stepfun, designed for general AI application workflows.

Step 1X is an image generation model from Stepfun, designed for image generation and creative visual workflows.

Open-weight vision-language MoE released May 29, 2026: 198B total / 11B active with ~256K context and an explicit reasoning mode, built for coding agents and search workflows.

~196B sparse-MoE open foundation model released February 2, 2026 (Apache 2.0), positioned as StepFun’s most capable open release for frontier reasoning and agents; base checkpoints also published.