Step 3.7 Flash

StepFun's Step 3.7 Flash release at the end of May 2026 under an Apache 2.0 license is drawing attention because it combines open weights, multimodal capability, and low active compute in a package developers can actually try locally. The Shanghai-based lab, backed by investors including Tencent, released the model to According to StepFun's Hugging Face model card as cited in coverage, Step 3.7 Flash supports a 256K context window, three reasoning levels, and throughput of up to 400 tokens per second. The sparse MoE vision-language system has a 196B-parameter language backbone and a 1.8B-parameter vision encoder, activating about 11

StepFun released Step 3.7 Flash on May 29, 2026, a 198-billion-parameter sparse Mixture-of-Experts vision-language model targeting agentic use cases, adding native vision input and improved tool-use reliability over Step 3.5 Flash. It pairs a 196B-parameter language backbone with a 1.8B-parameter vision encoder (ViT) f Key published specifications include a 256K-token context window, up to 400 tokens/sec throughput, and three configurable reasoning levels (low, medium, high). StepFun positions Step 3.7 Flash as infrastructure for agentic coding and search workflows rather than as a raw scale leader. The MoE architecture keeps inferen

Lambda's deployment documentation explicitly names Step 3.7 Flash, attributes it to StepFun AI, and confirms the Apache 2.0 release with a 198B-parameter vision-language architecture, 11B active parameters per token, 256K context, and low/medium/high reasoning modes. The language backbone carries over from Step 3.5 Fla Lambda's own vLLM-MTP benchmarks show aggregate throughput of 7,982 tok/s on 4× B200, 6,224 tok/s on 8× H100, and 2,370 tok/s on 8× A100 under an 8192-in/2048-out, 32-concurrent workload. Architecture specifics include 288 routed plus 1 shared expert with top-8 sigmoid routing, 45 layers, 4096 hidden size, and S3F1 hyb

Step 3.7 Flash is described as a high-efficiency large AI model aimed at production-level Agents, open-sourced by StepFun on May 29, 2026 under the Apache 2.0 license, with Step 3.5 Flash identified as its predecessor in the Flash series. It uses a sparse MoE architecture totaling 196B+1.8B (ViT) parameters and activat The model delivers native multimodal understanding and execution plus internet and visual search enhancement, along with tool invocation and orchestration, and reportedly scored 67.1% on ClawEval-1.1 and 49.5% on Toolathlon for multi-tool collaboration. Following release, domestic hardware vendors including Tianshu Zhi