Cursor Uses Kimi K2.5 Training Model Caught Red-Handed! Deletes Posts, Official Response Does Sharp U-Turn

AI Programming Tool Cursor Releases Composer 2, Blog Post Fails to Mention Base Source, Model ID “kimi-k2p5-rl-0317” Captured via API Request in Less Than Two Hours
The person responsible for Moon’s Dark Side pretraining publicly questions licensing issues, then the official account quickly changes tone to “congratulations,” while Cursor co-founder never admits.
(Background: Rakuten boldly announces “Japan’s Largest AI Model,” community discovers the core is DeepSeek V3)

Table of Contents

Toggle

• From Questioning to Deletion, Then Official Congratulations
• What Are the Licensing Terms for Kimi K2.5?
• Cursor Provides Clarification

Early morning of March 20, Cursor (parent company Anysphere, latest valuation $29.3 billion) released the new model Composer 2. The official blog states “Continuing pretraining on the base model, combined with reinforcement learning,” but never mentions whose base it is.

Less than two hours later, developer @fynnso on X captured a model ID: kimi-k2p5-rl-0317-s515-fast, meaning “Kimi K2.5 + RL.”

From Questioning to Deletion, Then Official Congratulations

Moon’s Dark Side pretraining lead Du Yulun posted on X, saying that after testing Composer 2’s tokenizer, they found it “completely identical to our Kimi tokenizer,” and “almost certain this is our model further trained,” directly questioning Cursor co-founder Michael Truell: “Why disrespect our license and not pay any fees?”

The tweet was later deleted.

Elon Musk replied under @fynnso’s post, “Yeah, it’s Kimi 2.5,” further amplifying the topic.

The tone quickly shifted; official Moon’s Dark Side account @Kimi_Moonshot changed from accusation to congratulations, praising Cursor for releasing Composer 2, “We are proud to see Kimi K2.5 providing the foundation.” They also clarified that Cursor accessed Kimi K2.5 via a commercial agreement with Fireworks AI, with licensing compliance guaranteed by Fireworks AI’s platform agreement.

In other words, it’s not unauthorized use, but the licensing channel was not disclosed in public announcements.

What Are the Licensing Terms for Kimi K2.5?

Kimi K2.5 uses a modified MIT license, which explicitly states that products with over 100 million active tokens per month or over $20 million in monthly revenue must prominently display “Kimi K2.5” in the user interface.

Given Cursor’s paying user scale, the revenue threshold is almost certainly triggered. The terms are not complicated; the issue is that this was completely omitted in the blog post.

Cursor Provides Clarification

After the incident, Cursor co-founder Aman Sanger and VP of Developer Education Lee Robinson explained:

• The team evaluated perplexity across multiple bases, and Kimi K2.5 “proved to be the strongest”
• They added further pretraining on top of the base, combined with four times larger-scale reinforcement learning
• Deployment was done via inference and RL samplers from Fireworks AI
• In the final model, about 1/4 of the compute came from the base, the remaining 3/4 from Cursor’s own training

Aman Sanger admitted that omitting mention of the Kimi base in the blog post was “a mistake,” and the next model will be clearly labeled from the start.

This is not Cursor’s first exposure. When Composer 1 was released in November 2025, the community found its tokenizer was identical to DeepSeek, and the model occasionally output Chinese during inference, with no explanation provided at the time.