Step 3.7 Flash

Startup Fortune reports that StepFun's Step 3.7 Flash, released at the end of May 2026 under Apache 2.0 by the Shanghai-based AI lab (backed by investors including Tencent), is drawing attention for pairing open weights with strong claimed benchmark performance and lower active compute. The 196B-parameter language back According to StepFun's Hugging Face model card, the model supports a 256k context window, three reasoning levels, and throughput of up to 400 tokens per second, claims that explain rapid local-inference community interest. StepFun is positioning Step 3.7 Flash as infrastructure for agents and production deployments rat

MarkTechPost reports that StepFun released Step 3.7 Flash on May 29, 2026, as a multimodal 198B-parameter sparse Mixture-of-Experts vision-language model targeting agentic use cases, with native vision input and improved tool-use reliability over Step 3.5 Flash. The architecture pairs a 196B language backbone with a se Key specifications include a 256k context window, throughput of up to 400 tokens/sec, three reasoning levels (low, medium, high), and Apache 2.0 licensing. The model is designed for production-level Agents, with design priorities balancing speed, cost, reliable execution, and complex task-handling capabilities rather t

BenchLM's aggregator page (data as of September 23, 2026) lists Step 3.7 Flash with a capability score of 41.3/100 (field median 50.4), ranking 109th of 196 tracked models. Reported metrics include a $0.20 input / $1.15 output price per million tokens, a blended price of $0.67, a speed of 196 tok/s (field median 91), a Category rankings include Agentic at rank 56 of 105 (percentile 47th, score 35.1 across 7 verified benchmarks, weighted 22%) and Coding at rank 78 of 135 (percentile 43rd, score 35.9 across 2 verified benchmarks, weighted 20%), with Multimodal at score 71.0 across 2 verified benchmarks. The page notes that Reasoning, K

Lambda's deployment page details hardware throughput benchmarks for Step 3.7 Flash using vLLM with Multi-Token Prediction speculative decoding, measured at 8192 in / 2048 out tokens and 32 concurrent requests. Performance figures include 7,982 tok/s total throughput on 4× NVIDIA B200 GPUs (TTFT 2,183 ms, ITL 52.7 ms), The page confirms Step 3.7 Flash as StepFun AI's 198-billion-parameter sparse MoE vision-language model released open-weight under Apache 2.0, activating ~11B parameters per token with a 256K context and three reasoning modes. The architecture pairs a 196B language backbone (carried over nearly unchanged from Step 3.5

Step 3.7 Flash is a high-efficiency AI model released and open-sourced by StepFun on May 29, 2026, designed for production-level Agents and positioned as a successor within StepFun's Flash series (predecessor: Step 3.5 Flash). It employs a Sparse Mixture-of-Experts architecture with 196B language parameters plus a 1.8B Key capabilities include native multimodal understanding and execution, internet and visual search enhancement, tool invocation and orchestration, a 256k context length, and three selectable inference levels (low, medium, high). The model supports both cloud and on-premises deployment, and scored 67.1% on ClawEval-1.1