I think composer has currently by far the best price to performance ratio for coding (not counting subsidized subscription cost by OpenAI and Anthropic).
It's based on Kimi K2, but I think it's fair to say, that their RL really sets it apart from the other open weight models.
I think composer has currently by far the best price to performance ratio for coding (not counting subsidized subscription cost by OpenAI and Anthropic). It's based on Kimi K2, but I think it's fair to say, that their RL really sets it apart from the other open weight models.
Training any large model at scale is hard, and Cursor has trained several including agentic ones. https://cursor.com/blog/composer