# StepFun > A Shanghai lab shipping efficient multimodal models, many under Apache 2.0. Canonical: https://www.subconscious.dev/providers/stepfun · By The Subconscious Team · Updated October 11, 2026 - Founded: 2023 - Example models: Step 5 Preview, Step 3.7 Flash - Website: https://www.stepfun.com ## Overview StepFun is a Shanghai AI lab known for efficient multimodal models, with a mix of proprietary API models and open-weight releases. In September 2026 it launched Step 5 Preview, its new flagship for agentic work: a 600B-total, 27B-active mixture-of-experts model with 1M context and text and image input, which Artificial Analysis scores level with Kimi K3. Its cheaper workhorse, Step 3.7 Flash, came out in May 2026 as a 198B mixture-of-experts vision-language model with only 11B active parameters. It has 256K context, selectable reasoning levels, tool use and structured outputs, and it ships under Apache 2.0. StepFun's own API prices it at $0.20 in and $1.15 out per million tokens, and OpenRouter carries it too. StepFun's research leans on model and system co-design to cut decoding cost. The earlier Step3, a 321B multimodal reasoning model with 38B active, used Multi-Matrix Factorization Attention and Attention-FFN Disaggregation to stay cheap on both flagship and low-end accelerators. The lab also builds speech and audio models, and its Step-2 line reached the trillion-parameter range. The platform exposes an OpenAI-compatible API, and the open weights run on vLLM and SGLang. ## Upsides - Low-cost multimodal models that read images and video with long context. - Permissive Apache 2.0 weights on key releases, which helps self-hosting. ## Core use cases - Vision and video understanding inside cost-sensitive agents. - Teams that want a small-active-parameter model to self-host cheaply. ## Downsides - Trails frontier models on hard multimodal reasoning benchmarks. - Thin Western distribution and support, with China-hosted first-party inference. ## At a glance | Attribute | Value | |---|---| | Model access | Open (Apache 2.0) and API models | | Flagship models | Step 5 Preview, Step 3.7 Flash | | Speed | ~128 tok/s on Step 3.7 Flash | | Price | $0.20 in, $1.15 out (Step 3.7 Flash) | | Customization | Open weights to fine-tune | | Deployment | First-party API, OpenRouter | | Long context | Up to 1M (Step 5 Preview); 256K on Step 3.7 Flash | ## FAQ ### What is StepFun? StepFun is a Shanghai AI lab known for efficient multimodal models, with a mix of proprietary API models and open-weight releases. In September 2026 it launched Step 5 Preview, its new flagship for agentic work: a 600B-total, 27B-active mixture-of-experts model with 1M context and text and image input, which Artificial Analysis scores level with Kimi K3. Its cheaper workhorse, Step 3.7 Flash, came out in May 2026 as a 198B mixture-of-experts vision-language model with only 11B active parameters. It has 256K context, selectable reasoning levels, tool use and structured outputs, and it ships under Apache 2.0. StepFun's own API prices it at $0.20 in and $1.15 out per million tokens, and OpenRouter carries it too. ### What is StepFun best for? Vision and video understanding inside cost-sensitive agents; Teams that want a small-active-parameter model to self-host cheaply. ### How much does StepFun cost? StepFun pricing at a glance: $0.20 in, $1.15 out (Step 3.7 Flash). Rates change often, so check StepFun's pricing page before committing. ### How much context does StepFun support? StepFun's long-context support: Up to 1M (Step 5 Preview); 256K on Step 3.7 Flash. ### What are the downsides of StepFun? Trails frontier models on hard multimodal reasoning benchmarks; Thin Western distribution and support, with China-hosted first-party inference. ### What are the best alternatives to StepFun? Common alternatives include Subconscious, OpenAI, Anthropic, Google Vertex AI, Amazon Bedrock. Each has a head-to-head comparison with StepFun on this site. ## Comparisons - [Subconscious vs StepFun](https://www.subconscious.dev/compare/subconscious-vs-stepfun.md) - [OpenAI vs StepFun](https://www.subconscious.dev/compare/openai-vs-stepfun.md) - [Anthropic vs StepFun](https://www.subconscious.dev/compare/anthropic-vs-stepfun.md) - [Google Vertex AI vs StepFun](https://www.subconscious.dev/compare/google-vertex-vs-stepfun.md) - [Amazon Bedrock vs StepFun](https://www.subconscious.dev/compare/aws-bedrock-vs-stepfun.md) - [Together AI vs StepFun](https://www.subconscious.dev/compare/together-ai-vs-stepfun.md) - [Fireworks AI vs StepFun](https://www.subconscious.dev/compare/fireworks-vs-stepfun.md) - [Baseten vs StepFun](https://www.subconscious.dev/compare/baseten-vs-stepfun.md) - [Groq vs StepFun](https://www.subconscious.dev/compare/groq-vs-stepfun.md) - [Cerebras vs StepFun](https://www.subconscious.dev/compare/cerebras-vs-stepfun.md) - [DeepInfra vs StepFun](https://www.subconscious.dev/compare/deepinfra-vs-stepfun.md) - [Hugging Face Inference Providers vs StepFun](https://www.subconscious.dev/compare/hugging-face-vs-stepfun.md) - [Modal vs StepFun](https://www.subconscious.dev/compare/modal-vs-stepfun.md) - [Cloudflare Workers AI vs StepFun](https://www.subconscious.dev/compare/cloudflare-workers-ai-vs-stepfun.md) - [xAI vs StepFun](https://www.subconscious.dev/compare/xai-vs-stepfun.md) - [Mistral AI vs StepFun](https://www.subconscious.dev/compare/mistral-ai-vs-stepfun.md) - [DeepSeek vs StepFun](https://www.subconscious.dev/compare/deepseek-vs-stepfun.md) - [Moonshot AI vs StepFun](https://www.subconscious.dev/compare/moonshot-ai-vs-stepfun.md) - [Z.ai vs StepFun](https://www.subconscious.dev/compare/z-ai-vs-stepfun.md) - [Alibaba Cloud vs StepFun](https://www.subconscious.dev/compare/alibaba-cloud-vs-stepfun.md) - [Meta vs StepFun](https://www.subconscious.dev/compare/meta-vs-stepfun.md) - [Cohere vs StepFun](https://www.subconscious.dev/compare/cohere-vs-stepfun.md) - [SambaNova vs StepFun](https://www.subconscious.dev/compare/sambanova-vs-stepfun.md) - [Nebius vs StepFun](https://www.subconscious.dev/compare/nebius-vs-stepfun.md) - [Crusoe vs StepFun](https://www.subconscious.dev/compare/crusoe-vs-stepfun.md) - [fal vs StepFun](https://www.subconscious.dev/compare/fal-vs-stepfun.md) - [Novita AI vs StepFun](https://www.subconscious.dev/compare/novita-ai-vs-stepfun.md) - [Venice vs StepFun](https://www.subconscious.dev/compare/venice-vs-stepfun.md) - [Parasail vs StepFun](https://www.subconscious.dev/compare/parasail-vs-stepfun.md) - [Inference.net vs StepFun](https://www.subconscious.dev/compare/inference-net-vs-stepfun.md) - [GMI Cloud vs StepFun](https://www.subconscious.dev/compare/gmi-cloud-vs-stepfun.md) - [Thinking Machines vs StepFun](https://www.subconscious.dev/compare/thinking-machines-vs-stepfun.md) - [Sail Research vs StepFun](https://www.subconscious.dev/compare/sail-research-vs-stepfun.md) - [Morph vs StepFun](https://www.subconscious.dev/compare/morph-vs-stepfun.md) - [Relace vs StepFun](https://www.subconscious.dev/compare/relace-vs-stepfun.md) - [TypeSafe AI vs StepFun](https://www.subconscious.dev/compare/typesafe-ai-vs-stepfun.md) - [StepFun vs Runware](https://www.subconscious.dev/compare/stepfun-vs-runware.md) - [StepFun vs StreamLake](https://www.subconscious.dev/compare/stepfun-vs-streamlake.md) - [StepFun vs Wafer](https://www.subconscious.dev/compare/stepfun-vs-wafer.md) - [StepFun vs RunInfra](https://www.subconscious.dev/compare/stepfun-vs-runinfra.md) - [StepFun vs Particle.AI](https://www.subconscious.dev/compare/stepfun-vs-particle-ai.md) - [StepFun vs Luminal](https://www.subconscious.dev/compare/stepfun-vs-luminal.md) - [StepFun vs Infron](https://www.subconscious.dev/compare/stepfun-vs-infron.md) - [StepFun vs Inception](https://www.subconscious.dev/compare/stepfun-vs-inception.md) - [StepFun vs Dolphin](https://www.subconscious.dev/compare/stepfun-vs-dolphin.md) ## Sources - [Step model family, LLM Reference](https://www.llmreference.com/model-family/step) - [Step3 on Hugging Face](https://huggingface.co/stepfun-ai/step3) - [MarkTechPost on Step 5 Preview](https://www.marktechpost.com/2026/09/20/stepfun-launches-step-5-preview/) Pricing and model lineups change often; figures are a snapshot.