[dur-randir]

Based on their own numbers, 8B seems decent, but 34B not worth it compared to general-purpose trained models even on specific tasks. Which is an interesting result.