Cursor's Composer 2 Is Built on Kimi K2.5, and They Didn't Tell Anyone

On March 19, 2026, Cursor launched Composer 2 to over a million daily active users. The official blog post described it as the product of “continued pretraining” and “reinforcement learning.” It scored 61.7% on Terminal-Bench 2.0, beating Claude Opus at 58.0%. The pricing was aggressive: $0.50 per million input tokens, an 86% drop from Composer 1.5.

Nowhere in the announcement did Cursor mention that the model was built on top of Kimi K2.5, an open-source coding model from Beijing-based Moonshot AI.

A developer found out anyway.

The Leaked Model ID

Less than 24 hours after launch, a developer named Fynn (@fynnso on X) was debugging Cursor API requests and noticed the actual model identifier passing through: accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast.

That string is a roadmap. anysphere is Cursor’s parent company. kimi-k2p5 is the base model, Kimi K2.5. rl means reinforcement learning was applied on top. 0317 likely refers to a March 17 training date. s515 is an internal version tag. fast indicates an optimized inference variant.

Fynn’s post on X hit 444,000 views within a day, kicking off a licensing dispute that played out almost entirely in public.

Moonshot AI’s Accusation

Yulun Du, head of pretraining at Moonshot AI (also known as Darkside of the Moon), posted a direct response. He said his team tested Composer 2’s tokenizer and found it “identical to our Kimi tokenizer.” He called the model “almost certainly the result of further fine-tuning of our model” and tagged Cursor co-founder Michael Truell directly, asking why Cursor wasn’t respecting the Kimi K2.5 license or paying fees.

A second Moonshot employee also posted a confirmation before deleting it. Du’s original post was later deleted too, as the situation moved toward resolution.

The License Problem

The Kimi K2.5 model ships under a Modified MIT License with one critical clause. If the software is used in a commercial product with either more than 100 million monthly active users or more than $20 million in monthly revenue, the product must prominently display “Kimi K2.5” in its user interface.

Cursor’s parent company Anysphere hit $2 billion in annualized recurring revenue as of February 2026. That works out to roughly $167 million per month, over eight times the $20 million threshold in the Kimi license. Cursor has more than a million daily active users.

Cursor’s UI displayed “Composer 2” with no mention of Kimi anywhere. The blog post attributed performance gains to Cursor’s own continued pretraining and reinforcement learning pipeline.

Cursor’s Response

Lee Robinson, Cursor’s VP of Developer Education (who goes by “leerob” in developer circles), responded publicly on Hacker News and X. His initial comment: “Yep, Composer 2 started from an open-source base! Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training.”

When pressed further, Robinson was more direct: “Since people really want me to say this: KIMI K2.5. Yes, that is the base.”

Robinson said Cursor believed it was following the license through its inference partner Fireworks AI and their commercial agreements. He acknowledged that omitting the base model from the blog post was “a miss” and committed to including base model attribution in future releases.

Cursor co-founder Aman Sanger added that the team evaluated multiple base models and Kimi K2.5 was “the strongest.” The additional pre-training and reinforcement learning were then done at 4x scale through Fireworks AI’s platform.

The Quick Reversal

Within hours, the tone shifted. Moonshot AI’s official @Kimi_Moonshot account posted a congratulatory message, saying they were “proud to see Kimi K2.5 providing the foundation” and confirming that Cursor’s access through Fireworks AI was “part of an authorized commercial collaboration, with license compliance ensured through Fireworks AI’s commercial agreements.”

The speed of this reversal raised eyebrows. One reading is that Moonshot’s leadership and its pretraining team weren’t initially aligned on whether a commercial deal existed. Du’s accusatory tweet and its subsequent deletion suggest that the internal communication at Moonshot may have lagged behind whatever deal Fireworks AI had brokered. Another reading is that the deal was struck after the controversy forced both sides to the table.

Either way, the public-facing outcome was tidy: authorized use, commercial agreement in place, no license violation.

The Pattern: Composer 1 Was Also Open-Source Under the Hood

This isn’t the first time Cursor’s “proprietary” model turned out to be a fine-tuned open-source base. When Composer 1 launched in November 2025, users noticed that the tokenizer matched DeepSeek’s open-source model. Some users even reported occasional Chinese-language output during inference, a telltale sign of a Chinese-developed base model leaking through.

HN user mohsen1 summarized the pattern bluntly: “Cursor Composer 1 was Qwen and this is Kimi. IDE is based on VSCode. The entire company is built on packaging open source and reselling it.”

That framing is reductive but not wrong directionally. Cursor takes open-source foundations (VS Code for the editor, open-weight models for the AI) and adds a layer of product polish, tooling integration, and specialized training on top. The question is whether that approach requires transparency about what’s underneath.

The Transparency Question

Simon Willison, the well-known developer and AI commentator, called out the core issue on his blog: deliberately hiding the base model is “disrespectful of the researchers who created that model.”

The Kimi K2.5 team at Moonshot AI spent months building and open-sourcing a model that Cursor then used as a foundation. The Modified MIT License exists specifically to ensure that kind of contribution gets acknowledged. Whether the compliance happened through Fireworks AI’s commercial terms or through UI attribution is a legal question. The ethical question, whether users and the broader community deserve to know what model they’re actually running, is simpler.

Cursor’s blog post described Composer 2’s performance as coming from “our first continued pretraining run” and “reinforcement learning.” Both are true, but they tell an incomplete story. The base model matters. A reinforcement learning pipeline applied to Kimi K2.5 produces a different result than the same pipeline applied to, say, Llama or Mistral. The choice of base is a real engineering decision, and omitting it makes the announcement misleading by omission.

What Cursor Users Should Know

If you’re using Cursor with Composer 2, here’s what the model actually is:

Base model: Kimi K2.5 by Moonshot AI (Beijing-based)
Training on top: Continued pretraining + reinforcement learning by Cursor/Anysphere
Inference: Hosted through Fireworks AI
Cursor’s claim: ~75% of compute came from their own training, ~25% from the base
Performance: 61.7% on Terminal-Bench 2.0, 73.7% on SWE-bench Multilingual
Pricing: $0.50/$2.50 per million input/output tokens (Standard), $1.50/$7.50 (Fast)

The model works well. Terminal-Bench scores put it ahead of Claude Opus and behind GPT-5. The pricing is cheaper than nearly everything else in its performance class. None of that is in dispute.

The dispute is about whether a company valued at $29.3 billion, generating $2 billion in annual revenue, should be upfront about the open-source models that power its flagship product. After this week, Cursor says they will be. Whether that holds for Composer 3 remains to be seen.

The Bigger Picture for AI Coding Tools

This incident illustrates a growing tension in the AI coding tool space. Companies like Cursor, Windsurf, and others are competing on product experience while relying on open-source or third-party models as their AI backbone. The competitive pressure to present these models as proprietary is real, especially when raising money at multi-billion dollar valuations.

The Kimi K2.5 episode also highlights the geopolitical dimension. Chinese open-source AI models are becoming foundational infrastructure for American tech products. Kimi K2.5 is developed by Moonshot AI in Beijing. DeepSeek, Qwen, and other Chinese models power significant chunks of the AI coding ecosystem. That reality sits awkwardly alongside ongoing US-China tensions over AI exports, chip restrictions, and technology transfer.

HN commenter MangoCoffee put it sharply: “Americans cry about Chinese tech theft, then Cursor uses open-source Chinese tech and silence follows.”

Whether you see irony or pragmatism in that depends on your perspective. What’s harder to argue with is that open-source AI models from Chinese labs are doing exactly what open source is supposed to do: letting anyone build on top of them, regardless of geography. The license terms just ask for attribution. That shouldn’t be a hard ask.

Sources: