Step 3.7 Flash

Startup Fortune covered Step 3.7 Flash at the end of May 2026, noting that the Shanghai-based StepFun lab — backed by investors including Tencent — released the model under an Apache 2.0 license and drew attention by combining open weights, multimodal capability, and low active compute in a package developers could act Quoting StepFun's Hugging Face model card, the coverage reports a 256k context window, three reasoning effort levels, throughput of up to 400 tokens per second, a 196-billion-parameter language backbone paired with a 1.8-billion-parameter vision encoder, and roughly 11 billion activated parameters per token. The piece

StepFun released Step 3.7 Flash on May 29, 2026, as a multimodal Mixture-of-Experts model targeting agentic use cases, building on Step 3.5 Flash with native vision input and improved tool-use reliability. The model is a 198-billion-parameter sparse MoE vision-language model, pairing a 196-billion-parameter language ba Key specifications reported include a 256k-token context window, throughput of up to 400 tokens per second, three selectable reasoning levels (low, medium, high), and Apache 2.0 licensing. The vision encoder runs as a separate 1.8B ViT module that injects image representations into the language backbone, enabling image

The Baidu encyclopedia entry documents Step 3.7 Flash as a high-efficiency large AI model for production-level Agents, released and open-sourced by StepFun on May 29, 2026, sitting in the Flash series alongside predecessor Step 3.5 Flash. It describes a Sparse MoE architecture with 196B+1.8B (ViT) total parameters and The entry lists three selectable inference levels (low, medium, high), supports both cloud and on-premises deployment, and cites benchmark scores of 67.1 on ClawEval-1.1 and 49.5 on Toolathlon for multi-tool collaboration. It also notes that companies including Tianshu Zhixin, Haiguang Information, and Biren Technology