Cursor launched 'Composer 2' without disclosing it was built on Moonshot AI's open-source Kimi K2.5, only admitting it after public exposure.
Cursor launched Composer 2, marketed as 'frontier-level coding intelligence,' without disclosing it was fine-tuned from Moonshot AI's open-source Kimi K2.5 model. An X user discovered the undisclosed base model via code identifiers. Cursor's VP acknowledged the base model, claiming ~75% of compute came from their own training. Cursor co-founder Aman Sanger admitted it was 'a miss' not to disclose this upfront. Kimi confirmed the usage was part of an authorized commercial partnership via Fireworks AI.
Cursor built Composer 2 on Kimi K2.5 with additional RL fine-tuning, meaning 75% of compute was Cursor's own training — a legitimate open-source strategy, not a scam. The bigger technical signal: Kimi K2.5 is now validated as a strong coding base model by a $29B company. If Cursor can squeeze frontier-level performance out of it, so can you.
Pull Kimi K2.5 from Hugging Face or via Fireworks AI's API this week and run it against your current coding benchmark — if it matches GPT-4o on your task at lower cost, you have a new default base model for fine-tuning.
Suggested exercise · not independently tested.
Go to fireworks.ai, find the Kimi K2.5 model endpoint, and run this prompt: 'Write a Python function that parses nested JSON with error handling and returns typed dataclasses.' Compare output quality and latency directly against your current model.
Tags
Sources