Step 3.7 Flash

Startup Fortune reports that Shanghai-based StepFun, backed by investors including Tencent, released Step 3.7 Flash at the end of May 2026 under an Apache 2.0 license as a sparse mixture-of-experts vision-language system. According to the article citing StepFun's Hugging Face model card, the model has a 196 billion par The piece also states that Step 3.7 Flash supports a 256K context window, three reasoning levels, and claimed throughput of up to 400 tokens per second, positioning the release as infrastructure for agent workflows and local deployment scenarios. It frames StepFun as a serious efficiency-focused competitor in the Chine

MarkTechPost reports that on May 29, 2026, StepFun released Step 3.7 Flash, a 198B-parameter sparse Mixture-of-Experts vision-language model designed for coding agents and search workflows. The model pairs a 196B-parameter language backbone with a 1.8B-parameter vision encoder (ViT), adding native vision input and impr The article frames Step 3.7 Flash as StepFun's entry into efficiency-focused, multimodal MoE models aimed at developers building agentic pipelines. It emphasizes coding-agent reliability and search-workflow integration as concrete target workloads, while positioning the release as a step up from the Step 3.5 Flash gene

The Baidu Baike entry documents Step 3.7 Flash as a StepFun model released and open-sourced on May 29, 2026 under the Apache 2.0 license, following its predecessor Step 3.5 Flash. It specifies a Sparse MoE architecture with 196B language parameters plus a 1.8B ViT vision encoder, activating roughly 11B parameters per t Reported benchmark scores include 67.1% on ClawEval-1.1 (daily autonomous task execution) and 49.5% on Toolathlon (multi-tool collaboration), with capabilities spanning native multimodal understanding, internet and visual search enhancement, and tool invocation/orchestration. The entry also notes that companies includi