StepFunStep 3
Step3-VL-10B Base
The Base release exposes Step3-VL-10B before its chat post-training. The family pairs a Perception Encoder with Qwen3-8B and uses fully unfrozen multimodal pretraining.
Research release
Keep post-training claims separate
The shared card describes the family’s full training pipeline. The Base checkpoint should not inherit the final chat model's reinforcement-learning or PaCoRe results.StepFun Step3-VL-10B-Base model card
Adapt perception and language together
Practical guidance. This checkpoint fits work on new document types or visual domains. Choose the chat model if the immediate need is an assistant that follows instructions. A base checkpoint requires a defined training and evaluation process.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | Step3-VL-10B-Base |
| License | Apache 2.0.StepFun Step3-VL-10B-Base model card |
| Published stages | The model table lists Base and Chat as separate downloads.StepFun Step3-VL-10B-Base model card |
Related models
- 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