Step 3
Step 3 is a 321B-parameter visual reasoning model with 38B active parameters. Its design combines Multi-Matrix Factorization Attention with Attention-FFN Disaggregation.
Separate model design from service access
The published weights preserve the MFA attention architecture. The FP8 download is a format of the same model. API retirement does not remove these published weights.StepFun step3 model cardStepFun model retirement and migration
Keep old integrations explicit
Practical guidance. Use the weights for local work or architecture research. An application that still sends step-3 to the hosted API needs a model change. Recheck image handling and tool behavior after that change.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | step3step3-fp8step-3 |
| Context window | 65,536 tokens Maximum context specified by the Step 3 weight card. StepFun step3 model card |
| License | Apache 2.0 for code and weights.StepFun step3 model card |
| Model composition | 316B language parameters within the 321B visual-language model.StepFun step3 model card |
| API retirement | The step-3 API stopped on 8 July 2026. StepFun recommends Step 3.7 Flash.StepFun model retirement and migration |
Related models
- Step 3.7 Flash
StepFun · Step 3
A visual reasoning model with a 196B language backbone and 1.8B vision encoder for images, video, code and tool calls.
Launch date not confirmedAPI · Downloadable - Step3-VL-10B
StepFun · Step 3
Step3-VL-10B combines a 1.8B PE-lang visual encoder with a Qwen3-8B decoder.
Launch date not confirmedDownloadable