INTELLECT-3
INTELLECT-3 adapts GLM-4.5-Air-Base through supervised fine-tuning and large-scale reinforcement learning. It targets reasoning, coding, and agent tasks.
What the release provides
Training uses prime-rl and the verifiers environment library. Prime publishes the training and evaluation environments through its Environments Hub, alongside model weights and the training framework.INTELLECT-3 official model card
How to use it well
Practical guidance. A sparse model still needs storage and memory for its full weights. Plan the deployment around total memory use, then measure speed with your intended sequence lengths.
FP8 weights are a deployment format
Prime also publishes INTELLECT-3-FP8. This is a reduced-precision package of INTELLECT-3 and shares this model page. Its model card gives a single-H200 serving example for FP8.INTELLECT-3 FP8 weight package
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Context window | 131,072 tokens Configured position limit. Serving limits depend on the runtime and settings. INTELLECT-3 model configuration |
| Repository license | MITINTELLECT-3 official model card |
| Parameters | 106 billion total; 12 billion activeINTELLECT-3 official model card |
| Base model | GLM-4.5-Air-BaseINTELLECT-3 official model card |
Related models
- INTELLECT-2
Prime Intellect · INTELLECT
A 32-billion-parameter reasoning model trained with distributed reinforcement learning.
Launch date not confirmedDownloadable - INTELLECT-3.1
Prime Intellect · INTELLECT
A further reinforcement-learning release of the INTELLECT-3 reasoning model.
Launch date not confirmedDownloadable