Qwen3.5-0.8B-Reverse-Text-RL
The model continues from Prime's Qwen3.5 reverse-text SFT checkpoint. It is intended as a fixed teacher or sampler for distillation and training-framework research.
Research release
What the release provides
Prime used 20 reinforcement-learning steps with one inference H200 and one training H200. The card reports a mismatch between training and inference distributions and learning that stalls after about 20 steps.Qwen3.5-0.8B-Reverse-Text-RL official model card
How to use it well
Practical guidance. Use the release to reproduce the documented training behavior. Do not treat its narrow reversal score as evidence of broad reasoning ability.
The release has a deliberately narrow purpose
Prime explicitly says this checkpoint is not intended for general use. Its short run and unresolved training mismatch are part of the research context.Qwen3.5-0.8B-Reverse-Text-RL official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Repository license | Apache 2.0Qwen3.5-0.8B-Reverse-Text-RL official model card |
Related models
- Qwen3.5-0.8B-Reverse-Text-SFT
Prime Intellect ยท Prime RL research models
A deliberately undertrained Qwen3.5 model for training-system research.
Launch date not confirmedResearch releaseDownloadable