Step3-VL-10B
Step3-VL-10B combines a 1.8B PE-lang visual encoder with a Qwen3-8B decoder. StepFun jointly trained the components on 1.2 trillion multimodal tokens before supervised and reinforcement learning.
Check the inference method behind a result
The card distinguishes sequential reasoning from PaCoRe. It also records a correction to comparison-model benchmark settings. A result from parallel reasoning includes extra inference work.StepFun Step3-VL-10B model card
Compare one attempt with parallel attempts
Practical guidance. For image extraction, begin with the normal inference path. Use parallel reasoning only when the task benefits from multiple interpretations. Keep both latency and the total generated tokens in your comparison.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | Step3-VL-10BStep3-VL-10B-FP8 |
| License | Apache 2.0.StepFun Step3-VL-10B model card |
| Parallel reasoning | The published PaCoRe mode combines 16 parallel attempts. Its reported settings use up to 128K context.StepFun Step3-VL-10B model card |
| Formats | The FP8 repository is a quantized form of the chat release.StepFun Step3-VL-10B model card |
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 PaCoRe 8B is a reasoning model trained to combine parallel solution paths.
Launch date not confirmedResearch releaseDownloadable- Step3-VL-10B Base
StepFun · Step 3
A visual-language base model combining a Perception Encoder with Qwen3-8B before chat post-training.
Launch date not confirmedResearch releaseDownloadable