Async RL seems to be the main difference in how this model was trained. Impressively they're open sourcing the training framework and weights.
However key information missing from the article:
- Benchmark comparisons against SOTA models of similar size
- Compute efficiency: No discussion of cost, power consumption, or efficiency metrics compared to other training approaches
- Training stability - They mention "rewards and evaluations continue to rise, and training remains stable" but don't discuss any instability challenges common in RL training. Would be interesting to see differences with their async approach
