On March 19, 2026, Cursor published a blog post announcing Composer 2. The benchmarks were real: 61.7% on Terminal-Bench 2.0, beating Claude Opus 4.6 (58.0%) while costing one-tenth the price. The launch was presented as a breakthrough from a 50-person team that had spent months training and refining an in-house coding model.
Forty-eight hours later, the developer community had a very different conversation.
The Discovery
It started with a developer debugging an API endpoint. While inspecting Cursor's OpenAI-compatible API responses, they found a model ID that shouldn't have been there:
accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast
The string is not subtle. kimi-k2p5 is a reference to Kimi K2.5, an open-weight model from Chinese AI lab Moonshot AI released two months earlier. The rl suffix indicates reinforcement learning fine-tuning — exactly the technique Cursor described as their own innovation in the launch post.
Elon Musk quote-tweeted the discovery with two words: "Yeah, it's Kimi 2.5."
That was enough. The story pivoted from celebrating a product launch to examining what had actually been built, and by whom.
What Cursor Actually Shipped
Cursor's own blog post, read carefully, was technically accurate but tonally misleading. The company described "continued pre-training of a base model combined with reinforcement learning" without specifying which base model. In isolation, that framing implied an in-house foundation. In context of the discovered model ID, it read differently.
The actual composition, as Aman Sanger (Cursor co-founder) later clarified on X:
- Base model: Kimi K2.5 (chosen after evaluating several options — it had the best perplexity scores)
- Added work: Continued pre-training + 4× scaled reinforcement learning
- Compute split: Roughly 25% from Kimi's base, 75% from Cursor's additional training
The result is genuinely Cursor's product in the same sense that a car is the assembler's product even when the engine comes from elsewhere. But the launch blog post never mentioned the engine supplier.
The Evidence That Closed the Case
Moonshot AI's head of pre-training, Yulun Du, ran a test that was hard to argue with. He fed samples of Composer 2's output through tokenizer analysis and confirmed: the tokenizer was byte-for-byte identical to Kimi's. When two unrelated models share the same tokenizer, it is vanishingly unlikely they are independent implementations.
Du tagged Cursor's co-founder directly: "Why aren't you respecting our license, or paying any fees?"
The model ID, the tokenizer match, and the timing — combined — made the technical case for dependency on Kimi K2.5 essentially incontrovertible.
The License Problem
This is where the story gets legally interesting, not just ethically interesting.
Kimi K2.5 is open-source under a Modified MIT License that Moonshot AI wrote with specific commercial conditions. The clause that matters:
Commercial products or services exceeding $20 million in monthly revenue must prominently display "Powered by Kimi K2.5" in the user interface.
Cursor's revenue figures, widely reported in 2026, put the company at roughly $2 billion annualized ARR — roughly 8× above the monthly threshold that triggers the attribution requirement.
Composer 2 launched on March 19 with no mention of Kimi K2.5 anywhere in the UI, the docs, or the blog post. That is the factual basis for what became a licensing controversy.
The Fireworks Resolution
Cursor accessed Kimi K2.5 through Fireworks AI, a hosted inference platform that provides commercial access to open-weight models under licensing agreements. This detail matters because it changes the nature of the dispute from "theft" to "attribution failure."
Moonshot AI's official statement, issued days after the controversy broke, read: "Congratulations to the Cursor team on Composer 2! We're proud that Kimi K2.5 provides the foundation."
This is not the language of a violated party demanding remedies. It reads more like a diplomatic acknowledgment of a partnership that wasn't communicated clearly. Moonshot AI further clarified that the Fireworks arrangement was fully compliant.
Lee Robinson, Cursor's VP of Developer Experience, acknowledged the disclosure failure directly: "Not mentioning Kimi as the base in the blog post was a mistake. We'll correct it in the next model."
The controversy had narrowed from "they stole Kimi" to "they didn't credit Kimi" — which is real but different in kind.
Kimi K2.5: The Model Worth Knowing About
The story has obscured something genuinely interesting about Kimi K2.5 itself. It is not a consolation prize.
Released in January 2026, Kimi K2.5 is an open-weight model with capabilities that explain why Cursor chose it as a foundation:
- 256K token context window — four times what most competitors offer
- Native multimodal architecture — processes text, vision, and visual specifications in a unified framework
- Agent Swarm — a parallel execution mode where the model can coordinate up to 100 autonomous agents working simultaneously on sub-tasks, trained using a Moonshot-developed technique called PARL (Parallel Agent Reinforcement Learning)
- Benchmark results — on Terminal-Bench 2.0: 61.7%. On BrowseComp (agent swarm mode): 78.4% versus 60.6% for standard agent execution
- Pricing — $0.60 per million input tokens via API, compared to $3–15 for comparable proprietary models
The Agent Swarm capability is particularly relevant to coding applications. A 256K context window can hold an entire mid-sized codebase. The swarm architecture can assign different agents to different modules simultaneously, reducing execution time on parallelizable tasks by up to 4.5×.
Kimi K2.5 is available on Hugging Face and can be run locally with quantization. For developers who want to experiment with the model Cursor built on top of, it is accessible today.
The Deeper Pattern: Chinese Open-Source as Global Infrastructure
This story is more interesting as a data point about the AI industry than as a scandal about one company's attribution failure.
The uncomfortable truth is that Chinese open-source AI models have quietly become foundational infrastructure for the global AI industry. Kimi K2.5, Qwen (Alibaba), DeepSeek, and GLM are not regional also-rans — they are competitive or superior to Western open-weight offerings at a fraction of the cost, and they are being used as base models by companies on every continent.
Cursor's case is not unusual. It is representative. A $50B-valued company chose an open-source Chinese model over closed proprietary alternatives because the open-source option was genuinely better. That is a market signal worth paying attention to.
The pattern also reveals something about where the competitive frontier in AI coding has shifted. Pre-training is becoming commoditized. The real differentiation is in fine-tuning methodology, RL pipelines, workflow integration, and developer experience. Cursor's moat was never the base model — it was the UX, the IDE integration, and the millions of hours of developer interaction data. Kimi K2.5 was the engine; Cursor built the car around it.
What This Means for Developers
If you use Cursor, the practical takeaway is straightforward: the tool is as good as it was before the controversy. The model quality is real. The attribution failure is a governance issue, not a capability issue. For a broader comparison of how Cursor stacks up against Claude Code and other AI coding tools, see our AI dev tool power rankings.
If you are evaluating AI models for your own products or workflows, this episode is worth studying beyond the headlines:
- Inspect what you buy. Model attribution is increasingly unreliable at face value. The technical markers — tokenizer signatures, model IDs in API responses — are verifiable. The commercial claims are not.
- Open-source licenses in AI are catching up. Kimi K2.5's Modified MIT is more sophisticated than most. As open-source AI models proliferate, more will carry commercial conditions. "Open-source" does not mean "unconditional free for commercial use."
- The real moat is integration, not foundation models. Cursor's competitive position rests on being deeply embedded in the developer workflow — not on owning the base model. For most organizations building on AI, the same logic applies.
- Kimi K2.5 is worth knowing about. Whether as a foundation for your own fine-tuning or as an API endpoint for development work, it is a capable model with a price-performance ratio that proprietary alternatives cannot match.
The Lesson for the Industry
Cursor recovered from this episode faster than most companies would. The acknowledgment was prompt, the technical explanation was detailed, and Moonshot AI chose to frame it as a successful partnership rather than a violation. The Fireworks intermediary gave both sides a compliant commercial relationship to point to.
But the underlying tension does not resolve this cleanly. As AI products become more layered — proprietary wrappers on open foundations, hosted models through intermediaries, fine-tuned derivatives marketed as original — the question of what "your model" means becomes genuinely ambiguous.
The industry is developing norms around disclosure in real time, driven by cases exactly like this one. The developers who understand what is actually running under the hood of their tools will be better positioned than those who accept marketing at face value.
The model ID in the API response was not supposed to be visible. But it was. And that visibility is now a permanent feature of the landscape.
