StepFun

Step 5 Preview

StepFun

Step 5 Preview is StepFun’s hosted flagship for agentic coding, research, professional knowledge work, and financial analysis. Its sparse mixture-of-experts architecture has 600B total parameters and 27B active per token. The documented API accepts text, images, and video, generates text, and supports a 1M-token context with up to 64K output. Tool calls, JSON Schema output, caching, and low, medium, or high reasoning effort support sustained workflows in an external application. It was publicly announced on September 20, 2026; September is used because earlier reported API availability dates are not resolved by a first-party dated changelog. Announced October 15 weights are future availability at this audit cutoff.

Step 3.7 Flash

StepFun

Step 3.7 Flash is an open-weight vision-language model for high-throughput agent execution. Its 198B total parameters combine a 196B sparse language backbone with a 1.8B vision encoder, activating about 11B parameters per token. It accepts images and text, supports a 256K context, and offers low, medium, and high reasoning levels. Training targets visual search, browser and terminal tool use, coding, and professional document workflows. API, web chat, and downloadable weights are available. Advisor-mode results depend on an additional larger model and should not be interpreted as standalone checkpoint performance.

Step 3.5 Flash

StepFun

Step 3.5 Flash is StepFun’s sparse language model optimized for reasoning, coding, and tool-driven agents. A 3:1 sliding-window/full-attention pattern reduces long-context inference cost while multi-token prediction increases throughput. The model activates 11B of 196B parameters and supports 256K context. Public access predates the February 12 technical-blog update; repositories and contemporaneous API catalogs establish the early-February release. The full weight footprint, context length, and selected reasoning mode affect deployment cost and latency.

Step3 VL 10B

StepFun

Step3 VL 10B is StepFun’s compact vision-language reasoning model, combining an approximately 1.8B visual encoder with a language backbone. It targets visual STEM reasoning and language generation from images. PaCoRe mode combines evidence from sixteen parallel rollouts, so its scores should be separated from ordinary single-model sampling. Apache-licensed base and post-trained checkpoints support research and local evaluation. The full weight footprint, context length, and selected reasoning mode affect deployment cost and latency.

Step-DeepResearch

StepFun

Step-DeepResearch is StepFun’s 32B language model post-trained for autonomous information gathering and professional report generation. It internalizes planning, search, reflection, cross-validation, and synthesis through agentic mid-training, supervised fine-tuning, and reinforcement learning. The technical report and reference harness were published on December 24, with application-based API access documented afterward. The model is separately identified as step-dr-1, rather than being only a generic research application. Public model weights and their license are not established by the open-source harness license.

Step 3

StepFun

Step 3 is StepFun’s open-weight multimodal reasoning model with a sparse language backbone and visual encoder. Multi-head latent attention reduces serving overhead while image-conditioned reasoning supports documents, charts, and general visual questions. The language backbone has 316B parameters; the full vision-language model has 321B. Apache-licensed weights and a technical report support private deployment and reproducible evaluation. The full weight footprint, context length, and selected reasoning mode affect deployment cost and latency.