Step 3.7 Flash

Step 3.7 Flash appears on the LLM Stats leaderboard ranked 127th overall, with capability-tier standing of average in general, below the top half in tool calling, legal, and finance categories, and coding/reasoning/math in the middle tier. Its blended price point is reported at approximately $0.25 per million tokens wi Benchmark evaluations sourced from StepFun's official scorecard include Terminal-Bench 2.1 (rank 34, score 0.59/1), SWE-Bench Pro (rank 36, score 0.56/1), and Humanity's Last Exam with tools in text-only mode (rank 8, score 0.47/1). The methodology notes indicate these are StepFun-internal or publisher-reported agentic

StepFun released Step 3.7 Flash at the end of May 2026 under an Apache 2.0 license, positioning it as an open-weights, inference-efficient alternative in the emerging "Flash" category of AI models. The model is a sparse mixture-of-experts vision-language system with a 196 billion parameter language backbone paired with According to StepFun's Hugging Face model card, Step 3.7 Flash supports a 256k context window, offers three selectable reasoning levels, and is claimed to deliver throughput of up to 400 tokens per second. The release is framed for practical local deployment and agent workflows, with the active-parameter figure highlig

StepFun released Step 3.7 Flash on May 29, 2026, as a 198B-parameter sparse Mixture-of-Experts vision-language model targeting agentic use cases. It pairs a 196B-parameter language backbone with a 1.8B-parameter vision encoder (ViT) for native image understanding, activating approximately 11B parameters per token durin Key specifications include a 256k-token context window, claimed throughput of up to 400 tokens per second, three reasoning levels (low, medium, high), and an Apache 2.0 license. The vision encoder runs as a separate 1.8B ViT module that injects image representations into the language backbone, enabling multimodal agent

Step 3.7 Flash was released and open-sourced by StepFun on May 29, 2026, as a high-efficiency large model in the Flash series succeeding Step 3.5 Flash. It targets production-level agent deployment, balancing speed, cost, reliable execution, and complex task handling as agent technology transitions from demos to enterp Technical specifications confirm a Sparse MoE architecture with 196B language parameters plus a 1.8B ViT vision encoder, activating only 11B parameters per token and reaching up to 400 tokens/s. It supports a 256k context length and three inference levels (low, medium, high), with both cloud and on-premises deployment