StepFun: Step 3.7 Flash

Step 3.7 Flash is a text generation model from StepFun, listed on api14 under the model id stepfun/step-3.7-flash. It is described by its provider as StepFun's latest high-efficiency multimodal Mixture-of-Experts model, pairing a 196B-parameter language backbone with a vision encoder and activating roughly 11B parameters. The provider release date is 2026-05-28. If you are choosing a model for a new project, the important framing is that this is a multimodal input model with text output, built around efficiency rather than raw scale.

On the input side, Step 3.7 Flash accepts text, images and video, and it produces text only. The context window is 262144 tokens, and the maximum output is 230400 tokens. Native image and video understanding comes from the vision encoder, so visual material can be passed in alongside text rather than being pre-converted to descriptions. Reasoning behaviour is exposed through the reasoning and include_reasoning parameters, and reasoning_effort lets you control how much effort is spent there. Structured outputs and response_format are supported, which matters when you need machine-readable results rather than prose.

Those capabilities point to a fairly specific set of use cases. Long documents plus long answers fit the 262144-token context and the 230400-token output ceiling: contract review, research synthesis, transcript analysis, codebase explanation. Because images and video are accepted as input, you can build pipelines that ask questions about screenshots, diagrams, slide decks, recorded demos or surveillance-style footage and get text back. Tool calling via tools and tool_choice makes it usable as the reasoning step in an agent loop, and structured_outputs plus response_format suit extraction tasks where the result feeds another system. The reasoning controls are useful when you want to trade depth against latency per request.

Calling it through api14 works the same way as any other OpenAI-compatible model. Point your client at the api14 base URL, use your api14 key, and set the model field to stepfun/step-3.7-flash. The supported request parameters are frequency_penalty, include_reasoning, logprobs, max_tokens, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs and top_p. Since the surface is OpenAI-compatible, existing SDKs and libraries generally work without changes beyond the model id and base URL.

There is no stated restriction in the facts beyond the modality split: input can be text, image or video, but output is text only, so anything requiring generated images, audio or video needs a different model. If your workload is pure text and does not benefit from the vision encoder or the reasoning controls, a simpler text-only model may be a better fit. Otherwise, the combination of a large context window, a large output ceiling and multimodal input makes this a reasonable default for document and media analysis tasks.