Step 3.7 Flash

In third-party benchmark rankings, Step 3.7 Flash holds a composite LLM Stats Score rank of 120, with a top-half placement in coding (99 of 267) and weaker showings in reasoning (117 of 363), math (150 of 327), and tool calling (141 of 194). The model ranks 32nd on Terminal-Bench 2.1 with a score of 0.59, 35th on SWE-B The aggregator notes that benchmark methodology specifics (harness, sampling, pass@k) are not disclosed by the publisher, so these figures should be treated as directional rather than definitive. Reported blended pricing is approximately $0.25 per million tokens, though this reflects marketplace data rather than StepFu

Step 3.7 Flash achieved top rankings on the Artificial Analysis benchmark in June 2026, securing first place in speed, cost-efficiency, and end-to-end performance while gaining significant traction on OpenRouter and Hugging Face. Output speed reached up to 416 tokens per second—one of the fastest among comparable model Multimodal understanding demonstrations showed the model identifying a dexterous robotic hand from its appearance, recognizing specific joint segments and fingertips, then autonomously searching for and compiling a comprehensive product specification table including manufacturer info, hardware configuration, load capac

StepFun's Step 3.7 Flash is a sparse mixture-of-experts vision-language model with a 196 billion parameter language backbone and a 1.8 billion parameter vision encoder, activating roughly 11 billion parameters per token. According to coverage citing the Hugging Face model card, the model was released at the end of May Reported throughput reaches up to 400 tokens per second, a figure that still requires independent validation across different hardware and workloads. The active-compute bill, rather than the headline parameter count, is framed as the key practical consideration for developers and startups evaluating local deployment. T

StepFun released Step 3.7 Flash, a multimodal Mixture-of-Experts vision-language model targeted at agentic coding and search workflows. According to MarkTechPost's coverage, the model pairs a 196B-parameter language backbone with a separate 1.8B-parameter ViT vision encoder for native image understanding, totaling 198B Step 3.7 Flash offers a 256k-token context window and up to 400 tokens/sec throughput, with the vision encoder injecting image representations into the language backbone rather than running end-to-end fused. The model is positioned for coding agents and search-style agentic use cases that benefit from the sparse-activa

Step 3.7 Flash is a high-efficiency, production-grade agent model released and open-sourced by StepFun on May 29, 2026 under the Apache 2.0 license. It uses a Sparse Mixture-of-Experts architecture with 196B total language parameters plus a 1.8B vision encoder, activating only about 11B parameters per token, and claims The model supports a 256K context window, native multimodal understanding, internet and visual search enhancement, and tool invocation/orchestration, and offers three selectable inference levels (low, medium, high) with both cloud and on-premises deployment. StepFun reports benchmark scores of 67.1% on ClawEval-1.1 (re

OpenRouter lists Step 3.7 Flash as StepFun's latest high-efficiency multimodal Mixture-of-Experts model, pairing a 196B-parameter language backbone with a vision encoder for native image and video understanding and activating roughly 11B parameters per token. The page documents a 262K (rendered as 262,144-token) contex OpenRouter routes Step 3.7 Flash across three providers — StepFun, DeepInfra, and NovitaAI — with identical listed pricing but materially different throughput and latency profiles: DeepInfra reports 0.12s P50 latency at 84 tok/s and 99.89% uptime, StepFun sits at 5.15s / 42 tok/s / 98.69%, and NovitaAI at 2.55s / 8 tok