Cursor's 'In-House Model' Was Someone Else's — Composer 2 and Kimi K2.5

The Name "Kimi K2.5" Was Left in Cursor Composer 2's API

On March 19, AI coding editor Cursor announced "Composer 2." It was billed as their most powerful coding model, developed in-house.

Less than 24 hours after the announcement, developer Fynn was poking around Cursor's OpenAI-compatible API and found a certain model ID:

Breaking down this ID reveals everything:

In other words, Composer 2 was based on "Kimi K2.5," an open-weight model published by Chinese AI company Moonshot AI, with additional reinforcement learning applied for coding tasks. The words "Kimi K2.5" appeared nowhere in Cursor's announcement blog post or release notes.

What Kind of AI Model Is Composer 2?

Cursor introduced Composer 2 as "their highest-performing coding model", emphasizing "continued pre-training" and "scaled reinforcement learning" as proprietary innovations.

Technically, it uses a method called "compaction-in-the-loop reinforcement learning." When generation reaches the token limit, the model compresses its own context to approximately 1,000 tokens before continuing. By incorporating this summarization into the learning loop, the model reportedly handles long codebases without losing coherence.

The technology itself is commendable. The problem was that they said nothing about where the foundation came from.

What Is Kimi K2.5, and Why Does It Matter?

Kimi K2.5 is an open-weight model developed by Chinese AI startup Moonshot AI. It handles coding, image recognition, and agentic behavior in a single model, and its "Agent Swarm" feature can run up to 100 instances simultaneously for parallel task processing.

Open-weight means the model's weights (parameters) are publicly available and anyone can download and use them. However, unlike "open source," there are license restrictions on how the model can be used.

And that license is at the heart of this controversy.

What Does the Kimi K2.5 License Say?

Kimi K2.5 is released under a "Modified MIT License" that states:

There are two key points here:

• 1. This obligation is explicitly stated to apply to derivative works. In other words, the defense that "we applied reinforcement learning so it's a different model" is unlikely to hold
• 2. Cursor's annual revenue is approximately $2 billion (ARR), which translates to roughly $167 million per month -- over 8 times the threshold

Yet Cursor's UI displays only "Composer 2." The words "Kimi K2.5" appear nowhere in blog posts or documentation.

How Did Moonshot AI Respond?

There were multiple developments from the Moonshot AI side:

• -- At least two Moonshot employees posted on social media claiming "this is a license violation," then later deleted the posts
• -- Pre-training lead Yulun Du publicly pointed out tokenizer similarities and raised questions about license compliance

The fact that posts were deleted suggests negotiations may be happening behind the scenes. However, given that someone at the leadership level has spoken publicly, this doesn't appear to be a simple misunderstanding.

Why Is Cursor Staying Silent?

As of this writing, Cursor has made no official comment.

There may be business reasons behind the silence. Cursor's parent company Anysphere has received a $29.3 billion valuation and is reportedly aiming to raise it to $50 billion. In those negotiations, the "in-house developed model" narrative would have been a crucial selling point.

Admitting "it was actually just another company's model with reinforcement learning applied" could fundamentally undermine that story.

Does "We Applied RL So It's a Different Model" Hold Up?

One possible argument Cursor could make is that "the RL modifications were substantial enough that it is no longer a derivative work."

However, Kimi K2.5's license explicitly states that attribution obligations extend to derivative works. Since the model weights were used as the starting point, no amount of RL can make the claim that it was "built from scratch."

In the HN comment thread, opinions like this appeared:

Technical achievement and attribution obligations are separate matters. If anything, producing outstanding results should have been all the more reason to show respect for the foundation. That appears to be the prevailing sentiment in the community.

Developer Community Reaction

The story generated a Hacker News thread with over 210 points and spread rapidly on X (formerly Twitter).

Reactions broadly fell into three camps:

Three scenarios are plausible going forward: add the attribution and settle, argue that the RL modifications constitute "sufficient transformation" and fight it, or rebuild Composer 2 on a different base model. Whichever path they choose, it will set a precedent in the AI coding tool industry.

Who Verifies What's Inside AI Coding Tools?

We pay $20 a month for AI tools and use them without knowing what model is running inside. Food products are required to list their ingredients, yet there is no such requirement for AI models.

In movie terms, it's like a ghostwriter penning a screenplay and a star director putting only their own name on the credits. Even if the work is excellent, stealing credit erodes trust across the industry.

How Cursor will respond remains to be seen. But one thing is certain: the model ID left in the API was like a ghostwriter who forgot to erase their name. And the internet didn't miss it.

References

• 1. Fynn (@fynnso) -- Tweet discovering the model ID
• 2. Awesome Agents -- Cursor's Composer 2 Is Kimi K2.5 With RL - And No Attribution
• 3. Hacker News -- Cursor Composer 2 is just Kimi K2.5 with RL
• 4. Phemex -- Moonshot AI Accuses Cursor of License Breach
• 5. The Decoder -- Cursor takes on OpenAI and Anthropic with Composer 2
• 6. Kimi K2.5 Tech Blog -- Visual Agentic Intelligence
• 7. Cursor Changelog -- Composer 2 Announcement
• 8. Mark Kretschmann (@mark_k) -- On the base model of Composer 2