A Single API String Exposed Cursor’s Secret: Composer 2 Runs on Moonshot AI’s Kimi K2.5

On March 19, a $29.3 billion coding startup launched what it called a breakthrough proprietary model. Within 24 hours, a developer found this string in the API response: kimi-k2p5-rl-0317-s515-fast. That one line of text unraveled the entire narrative.

Cursor’s Composer 2 — the model that supposedly beat Claude Opus 4.6 on coding benchmarks at one-tenth the price — wasn’t built from scratch. It was built on top of Kimi K2.5, an open-source model from Beijing-based Moonshot AI. And Cursor’s original blog post didn’t mention that at all.

The fallout has been swift, touching on everything from open-source licensing norms to the geopolitical anxieties of Western developer tools depending on Chinese foundation models. Here’s the full story.

How the Discovery Unfolded

When Cursor announced Composer 2, the messaging was clear: this was an “in-house” model representing Cursor’s “first continued pretraining run.” The blog highlighted benchmark scores — 61.3 on CursorBench, 73.7 on SWE-bench Multilingual — and an aggressive price point of $0.50 per million input tokens. No mention of any upstream model provider.

That story lasted less than a day.

A developer named Fynn spotted the model identifier kimi-k2p5-rl-0317-s515-fast in Cursor’s API configuration and posted it publicly. The identifier literally spells out the lineage: Kimi K2.5, plus reinforcement learning, dated March 17. Within hours, Moonshot AI’s pre-training lead Du Yulun ran his own analysis and confirmed that Composer 2’s tokenizer was “completely identical” to Kimi’s tokenizer. He directly questioned Cursor co-founder Michael Truell about license compliance.

Elon Musk weighed in with a characteristically blunt three-word reply: “Yeah, it’s Kimi 2.5.”

What Cursor Actually Admitted

Facing mounting evidence, Cursor’s leadership responded in stages.

First, Lee Robinson, Cursor’s vice president of developer education, acknowledged on X: “Yep, Composer 2 started from an open-source base!” He framed the contribution split as roughly 25/75 — about a quarter of the pretraining compute came from the Kimi K2.5 base, with Cursor doing the remaining three-quarters through fine-tuning and continued training.

Then co-founder Aman Sanger posted a more direct concession: “It was a miss to not mention the Kimi base in our blog from the start.”

The distinction Cursor is drawing — between using an open-source base model (which is legal and common) and claiming full credit for the result (which is a transparency problem) — is exactly where the controversy lives. Nobody disputes that fine-tuning open models is standard practice. The issue is that Cursor marketed Composer 2 as a proprietary breakthrough without disclosing its foundation.

The Kimi K2.5 Foundation

For those unfamiliar with the base model at the center of this story: Kimi K2.5 is a Mixture-of-Experts model with 1 trillion total parameters (32 billion active per request), released on January 27, 2026 by Moonshot AI. The company, also known as Darkside of the Moon, is backed by Alibaba and HongShan, and is one of China’s most prominent AI startups.

Kimi K2.5 was already making waves before the Cursor controversy. It was trained natively on 15 trillion mixed visual and text tokens, scored competitively against GPT-5.2 and Gemini 3 Pro on multiple benchmarks, and introduced an “Agent Swarm” capability allowing coordination of up to 100 parallel sub-agents.

The model is released under a Modified MIT License — open for commercial use, but with a significant condition: any entity generating over $20 million per month in revenue must prominently display “Kimi K2.5” as the underlying model. Given that Cursor has reportedly surpassed $2 billion in annualized revenue, this clause is directly relevant.

The Licensing and Geopolitical Dimension

The licensing question initially looked like it could escalate into a legal dispute. Du Yulun’s public challenge — “Why not respect our license and not pay any fees” — suggested Moonshot AI was prepared to push the issue.

But the resolution came surprisingly quickly. Moonshot AI’s official account posted a statement congratulating Cursor on the Composer 2 release, clarifying that Cursor’s access to Kimi K2.5 was through Fireworks AI, a U.S.-based platform for hosting and running open-source models, “as part of an authorized commercial partnership.” License compliance, according to both parties, was handled through Fireworks AI’s commercial agreements.

The legal question may be settled, but the geopolitical optics remain uncomfortable for some observers. A $29.3 billion American coding startup — used by developers at OpenAI, Uber, Spotify, and Instacart — built its flagship model on top of Chinese intellectual property. In the current U.S.-China AI competition climate, that fact alone is generating discussion in policy circles and on Hacker News threads.

It’s worth noting this isn’t a one-way dependency. Chinese AI labs routinely build on Western research, and open-source models are designed to be used globally. But the lack of upfront disclosure turned what could have been a positive story about open-source collaboration into a trust issue.

Composer 2 Performance: Good, but Context Matters

Strip away the controversy and look at the numbers alone, and Composer 2 is a solid product:

Benchmark	Composer 2	Claude Opus 4.6	GPT-5.4 Thinking
CursorBench	61.3	58.2	63.9
Terminal-Bench 2.0	61.7	58.0	75.1

The pricing is where Composer 2 genuinely stands out: $0.50/$2.50 per million tokens (input/output) versus competitors charging $2.50+/$15+ for similar-tier models. That’s an 86% cost reduction compared to Cursor’s own Composer 1.5.

But here’s the reframing that several analysts have pointed out: Cursor could have positioned these results as a win for open-source fine-tuning. A story about “we took the best open-source coding model, applied our proprietary RL training, and matched frontier models at a fraction of the cost” would have been genuinely compelling. Instead, by obscuring the Kimi base, Cursor created a narrative they then had to walk back.

What This Means for the AI Coding Tool Market

The Cursor-Kimi episode is a stress test for the entire AI coding ecosystem. Several dynamics are worth watching:

Model supply chain transparency is now a competitive issue. Developers are increasingly asking what’s actually under the hood of their AI tools. After this incident, expect competitors like Windsurf and Cline to be more explicit about which models power their features — and to use that transparency as a differentiator.

The “wrapper” critique gets sharper. Cursor has been fighting the perception that it’s a VS Code wrapper around someone else’s AI. Building Composer 2 was partly about proving otherwise. The revelation that Composer 2 itself wraps an open-source model — even if substantially fine-tuned — reinforces exactly the narrative Cursor was trying to escape. Musk’s xAI, which recently poached several of Cursor’s top engineers, may find this moment strategically useful.

Open-source model licensing needs clearer enforcement. The Kimi K2.5 Modified MIT License’s $20M/month revenue threshold is innovative, but the fact that compliance had to be publicly disputed before being confirmed suggests the enforcement mechanisms aren’t mature enough for the scale at which these models are being deployed.

The line between “fine-tuned” and “proprietary” is blurry. If 75% of the compute comes from your own training, is it still the base model? Cursor says no. Many developers say yes. This debate isn’t going away, and it affects how every AI company markets its models going forward.

FAQ

Is Cursor’s Composer 2 just a rebranded Kimi K2.5?

Not exactly. Cursor applied significant additional training — roughly 75% of the total compute, according to the company — through reinforcement learning and continued pretraining on top of the Kimi K2.5 base. The resulting model performs differently from stock Kimi K2.5 on coding benchmarks. However, the tokenizer is identical, and the base architecture is unchanged.

Did Cursor violate Kimi K2.5’s license?

Both Moonshot AI and Cursor say no. Cursor accessed Kimi K2.5 through Fireworks AI as part of an authorized commercial partnership. However, the initial lack of attribution in Cursor’s launch blog did raise questions about compliance with the license requirement to “prominently display” the Kimi K2.5 name for high-revenue commercial users.

How does Composer 2 compare to Claude Opus 4.6 and GPT-5.4 for coding?

Composer 2 scores 61.3 on CursorBench (vs. Opus 4.6’s 58.2) and 61.7 on Terminal-Bench 2.0 (vs. Opus 4.6’s 58.0). However, GPT-5.4 Thinking still leads at 63.9 and 75.1 on those same benchmarks respectively. The biggest differentiator is price: Composer 2 costs $0.50 per million input tokens, compared to $2.50+ for comparable models.

What is Moonshot AI, and who backs them?

Moonshot AI (also known as Darkside of the Moon) is a Beijing-based AI company that builds the Kimi model series. It’s backed by Alibaba and HongShan (formerly Sequoia China) and has become one of China’s most prominent AI startups. Kimi K2.5, released in January 2026, is their flagship open-source model with 1 trillion parameters.

Will this affect Cursor’s business or valuation?

Short-term impact appears limited — Cursor’s product still works well and the pricing advantage is real. But the trust issue could matter longer-term, especially as enterprise customers evaluate AI supply chain risks. Cursor’s $29.3 billion valuation was based partly on the premise of building proprietary AI capabilities, and this episode complicates that narrative.

Top AI Product

Leave a comment Cancel reply