PaCoRe 8B
PaCoRe 8B is a reasoning model trained to combine parallel solution paths. StepFun uses outcome-based reinforcement learning so the model can summarize and reconcile intermediate results.
Research release
Coordinate several attempts
The inference pipeline passes short messages between rounds of parallel reasoning. The paper's roughly two-million-token experiment describes total inference computation across trajectories, not a two-million-token input window.StepFun PaCoRe-8B model card
Compare the whole reasoning run
Practical guidance. Measure total generated tokens and elapsed time across all branches. A small parameter count can still produce a costly run when many branches execute. Compare it with a single-pass answer on the same problem and completion budget.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | PaCoRe-8B |
| Ancestry | The project starts from RLVR-8B-0926, itself post-trained from Qwen3-8B-Base.StepFun PaCoRe-8B model card |
| License | MIT.StepFun PaCoRe-8B model card |
| Release scope | Low, medium and high are inference configurations of one checkpoint.StepFun PaCoRe-8B model card |
Related models
- RLVR 8B 0926
StepFun · PaCoRe
A Qwen3-8B-Base derivative with reasoning-oriented post-training, used as the initial PaCoRe checkpoint.
Launch date not confirmedResearch releaseDownloadable - 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