Table of Contents
ToggleIn the early hours of March 20th, Cursor (parent company Anysphere, latest valuation of $29.3 billion) released its new generation model, Composer 2. The official blog stated that it was "the first time that the pedestal model was further pre-trained and then combined with reinforcement learning," but it never mentioned who the pedestal model was.
In less than two hours, developer @fynnso on X retrieved a model ID while debugging a Cursor API request: kimi-k2p5-rl-0317-s515-fast , which means "Kimi K2.5 + RL".
From questioning to deleting tweets, and then to official congratulations
Du Yulun, the head of pre-training for Lunar Dark Side, immediately posted on X, stating that after testing Composer 2's tokenizer, the team found it to be "completely identical to our Kimi tokenizer," and that "it's almost certain that this is the result of our model being further trained." He then directly questioned Cursor co-founder Michael Truell: "Why don't you respect our license and pay any fees?"
The tweet was subsequently deleted.
Elon Musk replied to @fynnso's post with "Yeah, it's Kimi 2.5", further amplifying the topic's popularity.
The tide quickly turned, with the official Dark Side of the Moon account, @Kimi_Moonshot, changing its tune from accusation to congratulations, congratulating the Cursor team on the release of Composer 2, stating, "We are proud to see Kimi K2.5 providing the foundation." The announcement also clarified that Cursor accesses Kimi K2.5 through a commercial agreement with Fireworks AI , and that license compliance is guaranteed by Fireworks AI's platform agreement.
In other words, this is not unauthorized use, but rather that the authorized channels were not disclosed in the public announcement.
What does the licensing agreement for Kimi K2.5 say?
Kimi K2.5 uses a modified MIT license, which explicitly stipulates that commercial products with more than 100 million monthly active tokens or more than $20 million in monthly revenue must prominently display "Kimi K2.5" in their user interface.
Given Cursor's scale of paying users, the monthly revenue threshold is almost inevitable. The terms themselves are not complicated; the problem lies in the fact that this was completely omitted when the blog was published.
Cursor (additional explanation)
Following the escalation of the incident, Cursor co-founder Aman Sanger and VP of Developer Education Lee Robinson both issued statements:
- The team conducted confusion tests on multiple bases, and the Kimi K2.5 "proved to be the strongest".
- On top of the base, further pre-training is performed, along with reinforcement learning with four times the computing power.
- Deployment via Fireworks AI inference and RL samplers
- In the final model, approximately one-quarter of the computing power comes from the base, while the remaining three-quarters comes from training the Cursor itself.
Aman Sanger admitted that it was a mistake not to mention the Kimi base when the blog post was published, and that the next model would be mentioned as soon as possible.
This isn't the first time Cursor has been caught out. When Composer 1 was released in November 2025, the community discovered that its tokenizer was identical to DeepSeek's , and the model would occasionally output Chinese characters during inference, without any explanation at the time.
Related reports
