StepFun releases Step 3.7 Flash: 198B MoE with 11B active parameters

StepFun has released Step 3.7 Flash, a sparse MoE model with 198B total parameters and roughly 11B active per token (8 of 288 experts). It claims up to 400 tokens/s, a 1.8B ViT encoder for native vision, and an Advisor Mode that scores 76.3% on SWE-Bench Verified at $0.19 per task versus 78.7% and $1.76 for Claude Opus 4.6.

• 198B total parameters, 11B active per token, 8 of 288 experts
• Up to 400 tokens/s plus a 1.8B ViT encoder for images
• SWE-Bench Verified: 76.3% at $0.19 per task vs $1.76 for Opus 4.6
• Terminal-bench 2.1: 59.5 versus 82.7 for GPT 5.5