Prime IntellectPrime RL research models
glm4-moe-tiny
This approximately 543-million-parameter model is trained on reversed-text examples. Its purpose is to provide a useful output distribution for reinforcement-learning framework checks.
Research release
What the release provides
The architecture has 24 layers, eight experts, and four active experts. Prime marks it for testing only and directs users to prime-rl. It is not the full GLM-4 model.glm4-moe-tiny official model card
How to use it well
Practical guidance. Use the small architecture to study expert routing and training behavior with limited compute. Do not transfer its results directly to a full-size GLM deployment.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Repository license | Apache 2.0glm4-moe-tiny official model card |
| Parameters | Approximately 543 millionglm4-moe-tiny official model card |
| Experts | 8 total; 4 activeglm4-moe-tiny official model card |
Related models
- qwen3-moe-tiny
Prime Intellect ยท Prime RL research models
A small Qwen3-style mixture-of-experts model for training-system research.
Launch date not confirmedResearch releaseDownloadable