At 2 a.m. on March 20, the AI programming tool Cursor (parent company Anysphere, latest valuation $29.3 billion) released its self-developed model Composer 2. The blog post stated that the performance improvement came from "the first continued pre-training of the base model, combined with reinforcement learning," without mentioning the source of the base model.
In less than two hours, developer @fynnso intercepted the actual model ID of Composer 2 while debugging a Cursor API request: `kimi-k2p5-rl-0317-s515-fast`, literally "Kimi K2.5 + RL". Du Yulun, the head of pre-training at Moonshot AI, subsequently tweeted, stating that after testing Composer 2's tokenizer, the team found it "completely identical to our Kimi tokenizer," and was "almost certain this is the result of our model being further post-trained." He directly questioned Cursor co-founder Michael Truell: "Why disrespect our license and not pay any fees?"
The tweet was later deleted. The controversy quickly escalated on social media, with Elon Musk replying "Yeah, it's Kimi 2.5" under @fynnso's post, further amplifying the topic's热度 (heat/hotness).
Kimi K2.5 uses a modified MIT license, which explicitly states: commercial products with over 100 million monthly active users or monthly revenue exceeding $20 million must prominently display "Kimi K2.5" in the user interface. Given Cursor's valuation and paid user base, the monthly revenue threshold was almost certainly triggered.
Subsequently, the风向 (wind direction/narrative) reversed. The official Moonshot AI account @Kimi_Moonshot posted early this morning, its tone shifting from accusation to congratulations: "Congratulations to the Cursor team on the release of Composer 2. 'We are proud to see Kimi K2.5 providing the foundation.'" The statement also clarified that Cursor accessed Kimi K2.5 through the RL and inference platform hosted by Fireworks AI, which constitutes an authorized commercial cooperation, and license compliance is guaranteed by Fireworks AI's commercial agreement.
Following Kimi's official statement, Cursor co-founder Aman Sanger and VP of Developer Education Lee Robinson followed up. Sanger explained the technical choice: the team evaluated multiple base models for perplexity, and Kimi K2.5 "proved to be the strongest." They then叠加 (superimposed/added) continued pre-training and high-compute reinforcement learning at 4x the size, deploying it through Fireworks AI's inference and RL samplers.
Robinson added that the final model's compute contribution from the base model was about 1/4, with the remaining 3/4 coming from Cursor's own training. Both acknowledged that failing to mention the Kimi base in the blog post "was a mistake" and stated that the next model would注明 (indicate/note) it promptly.
This is already the second time Cursor has been found using a Chinese open-source model without disclosure. In November 2025, when Composer 1 was released, the community discovered its tokenizer was identical to DeepSeek's, and the model occasionally output Chinese during inference. Cursor同样未作说明 (similarly did not provide an explanation at the time).
The discussion sparked by this incident has gone beyond license compliance itself. Hugging Face co-founder and CEO Clément Delangue commented that this is yet another validation of Chinese open-source, "Today, Chinese open-source is the biggest force shaping the global AI tech stack." He noted that the cutting-edge competition is no longer just about who trains from scratch, but about who adapts, fine-tunes, and productizes the fastest.
A noteworthy temporal coincidence: On March 15, Bloomberg reported that Moonshot AI was seeking up to $1 billion in a new funding round, with a valuation of approximately $18 billion, more than quadrupling in just three months, with Alibaba and Tencent both participating. Just five days later, the world's highest-valued AI programming tool was discovered to be based on Kimi K2.5. Anysphere, valued at $29.3 billion, evaluated and determined Kimi K2.5 to be the "strongest base," building its core product upon it. This might be the most direct market endorsement of Moonshot AI's technical capabilities.
At this juncture, before this funding round is even completed, the Cursor incident effectively served as a capability demonstration of Kimi for global developers. Whether the $18 billion valuation still undervalues Moonshot AI may need to be重新审视 (re-examined/reassessed).
