A Prime Intellect nyílt forráskódúvá tette 106 milliárd paraméteres Reinforcement Learning modelljét
Prime Intellect has released INTELLECT-3, a massive 106-billion-parameter model designed to handle complex math, coding, and reasoning through reinforcement learning. In a major move toward democratizing high-level model training, the company has open-sourced not just the model weights, but also the training framework and a massive collection of environments.
- INTELLECT-3 features 106 billion parameters trained using reinforcement learning (RL).
- Optimized for complex mathematical problems, coding tasks, and advanced reasoning.
- The release includes the open-source training framework (Prime-RL).
- Provides 500+ environments so developers can train their own models on the same stack.
Miért fontos?
Open-sourcing the entire stack—weights, framework, and environments—allows independent researchers and developers to train large-scale RL models that were previously limited to major labs.
Kapcsolódó hírek
Az Ant Group kiadta a LLaDA2.1 szerkeszthető diffúziós nyelvmodelleket
2026. február 23.