Prime IntellectPrime RL research models
qwen3-moe-tiny
This approximately 670-million-parameter model is fine-tuned on Prime's Reverse-Text-SFT data. It is a training-framework research model, not a general Qwen assistant.
Research release
What the release provides
The model has 24 layers and 16 experts, with four active experts. Its small trained output distribution supports reinforcement-learning checks in prime-rl. The card limits the intended use to testing.qwen3-moe-tiny official model card
How to use it well
Practical guidance. Compare the same training operation across this model and the GLM-style small model. Keep architectural differences visible when interpreting memory use and speed.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Repository license | Apache 2.0qwen3-moe-tiny official model card |
| Parameters | Approximately 670 millionqwen3-moe-tiny official model card |
| Experts | 16 total; 4 activeqwen3-moe-tiny official model card |
Related models
- glm4-moe-tiny
Prime Intellect ยท Prime RL research models
A small GLM-style mixture-of-experts model for training-system research.
Launch date not confirmedResearch releaseDownloadable