സ്റ്റെപ്പ് 3.7 ഫ്ലാഷ്

StepFun's sparse Mixture-of-Experts vision-language model, pairing a 196B-parameter language backbone with a vision encoder for native image and video understanding while activating roughly 11B parameters per token. It is engineered for high-frequency production workloads and reaches up to 400 tokens per second. StepFun built it for production agents spanning agentic, coding, search and multimodal work, with three selectable reasoning levels to trade speed and cost against depth.

StepFun organises the release around four claims rather than one headline: multimodal understanding that turns into action, web and visual search that reaches further, tool use and orchestration that stays coherent however long a run gets, and compatibility with the agent harnesses teams already use. The last one is not filler. The lab publishes a per-harness breakdown of its own coding benchmark across six popular agent frameworks, where the average rises from 56.5% for Step 3.5 Flash to 67.1%, and, the part it actually argues for, the spread between the best and worst harness narrows sharply. A model that only performs well on one scaffold is not usable in a production stack that runs several.

The efficiency case is made in cost rather than only in speed. StepFun reports 76.5% on SWE-bench Verified, 56.3% on SWE-bench Pro (five points up on the previous Flash generation) and 59.6% on Terminal-Bench 2.1 (up 6.1). It also ships an advisor arrangement in which this model drives the whole trajectory - calling tools, reading results, iterating - and consults a larger model only at the few inflection points where its own judgement falls short, such as planning or recovery from repeated failure. With that on, the lab measures 76.3% on SWE-bench Verified at 19 cents per agentic task against Claude Opus 4.6 at 78.7% for $1.76, or roughly 97% of the score for about a ninth of the cost.

Search is treated as part of reasoning rather than as an add-on, on the explicit reasoning that a model this size cannot hold the world in its weights and should instead be good at going to get things. StepFun reports 47.2% on Humanity's Last Exam with tools against 35.7% for Step 3.5 Flash, 75.8% on BrowseComp, 92.8% F1 on DeepSearchQA and 71.7% on ResearchRubrics, which it places ahead of GPT 5.5 at 61.5% and near Claude Opus 4.7 at 73.9%.

Vision follows the same logic. Rather than growing parametric knowledge, the model is trained to reach for visual tools: with a visual search tool it reports 79.2% on SimpleVQA, matching or beating models several times its size on long-tail entity recognition, and given a code tool that lets it crop, zoom and draw on an image it reports 95.3% on the V-star fine-grained perception set and 89.1% on the 4K high-resolution variant. StepFun extended it to phone interface operation as well, at 61.9% on an Android daily-task benchmark. The behaviour it found most interesting was one nobody trained for: after writing frontend code the model opened the interface, exercised the page it had just built and revised its own code from what it saw.

StepFun's own comparison table is where the trade is visible. Its 59.6% on Terminal-Bench 2.1 sits well behind Gemini 3.5 Flash at 76.2% and GPT 5.5 at 82.7%, and on professional-deliverable scoring it is behind the frontier cohort too. That is the Flash bargain stated plainly: agent reliability, perception and search at a rate that lets you run many concurrent agents, not the best score on any single hard benchmark. The weights are published in several precisions and run under vLLM, SGLang, transformers and llama.cpp, small enough to fit a single high-memory workstation, so this is a model you could host yourself; the copy served here is StepFun's.