Cursor's 'Frontier AI' Secret: Composer 2 Was Built on Kimi K2.5 — Here's What That Means

March 30, 2026  ·  Happycapy Guide

What Happened: The Cursor Composer 2 Incident

On March 19, 2026, Cursor launched Composer 2 — described in their announcement as “frontier-level coding intelligence that completes complex, multi-file tasks with one instruction.” The launch post emphasized performance benchmarks and a lower cost per task compared to competing coding agents. What it did not mention: the model powering it was Kimi K2.5, an open-weight AI released in January 2026 by Moonshot AI, an 80-person startup in Beijing.

Within hours, a developer posting on X noticed something in Cursor's API responses: the internal model identifier referenced a Moonshot AI endpoint. Screenshots spread quickly. By the next morning, the post had accumulated tens of thousands of reposts and Cursor was trending across AI communities. Cursor co-founder Aman Sanger issued a response the following day, acknowledging the attribution oversight and confirming the partnership with Moonshot AI.

What Is Kimi K2.5?

Moonshot AI released Kimi K2.5 on January 27, 2026. It is a Mixture-of-Experts (MoE) model with 1 trillion total parameters and approximately 32 billion active parameters per inference — a design that delivers high quality output while using far less compute than a fully dense model of the same scale.

Three capabilities distinguish Kimi K2.5 from comparable open-weight models:

At launch, Kimi K2.5 ranked 4th globally on coding benchmarks — behind Claude Opus 4.5, GPT-5.2, and Gemini 3 Pro. On HumanEval, it scores comparably to top frontier models at a fraction of the API cost. Moonshot AI is backed by Alibaba and HongShan (formerly Sequoia China) and is valued at $18 billion despite having fewer than 100 employees.

The Timeline

Why This Matters Beyond Cursor

The controversy is not primarily about Cursor — it is about a structural shift in how AI products are built. More and more developer tools are shipping model-agnostic interfaces on top of whatever open-weight model is cheapest, fastest, or highest-quality for that release cycle. The underlying model changes without the end user knowing.

• Chinese open-source AI is frontier-tier. Kimi K2.5 ranking 4th globally in coding — above many paid Western models — confirms that the open-source leadership gap is gone. Moonshot AI, Alibaba (Qwen), and DeepSeek are now competitive with GPT and Claude on specific benchmarks.
• Attribution standards are catching up. Several AI companies have released modified open-source licenses in 2026 requiring disclosure in commercial products. The Cursor incident will accelerate enforcement expectations across the industry.
• Users cannot assume model stability. If the AI tool you pay for can swap the underlying model without notice, the quality and behavior you depend on can change without warning. This makes model transparency a real product differentiator.
• Agent Swarm at 100 agents is a capability signal. Coordinating 100 parallel sub-agents in a single session is not something most consumer AI tools expose. Kimi K2.5 built it into the base model — a sign that multi-agent orchestration is becoming table-stakes infrastructure.

Coding AI Tools: Transparency Comparison

Frequently Asked Questions

What is Kimi K2.5 and who made it?

Kimi K2.5 is a 1 trillion parameter open-weight Mixture-of-Experts model released January 27, 2026, by Moonshot AI — an 80-person Beijing startup valued at $18 billion and backed by Alibaba and HongShan. It supports text, image, and video natively and includes Agent Swarm technology that coordinates up to 100 specialized agents simultaneously.

What was the Cursor Composer 2 controversy?

Cursor launched Composer 2 on March 19, 2026, as a new coding agent without disclosing that it was built on Kimi K2.5. A developer identified the Moonshot AI model identifier in Cursor's API responses within hours. The revelation spread virally. Cursor co-founder Aman Sanger acknowledged the attribution oversight and confirmed the Moonshot AI partnership the following day.

Is Kimi K2.5 better than Claude or ChatGPT for coding?

On coding benchmarks at launch, Kimi K2.5 ranked 4th globally — behind Claude Opus 4.5, GPT-5.2, and Gemini 3 Pro. Its HumanEval score is comparable to top frontier models at a significantly lower API cost. The Cursor Composer 2 product launch confirmed that at least one major commercial coding platform judged its quality sufficient to build a paid offering on.

What does the Kimi K2.5 license allow?

Kimi K2.5 uses a Modified MIT License. Personal and low-revenue use is unrestricted. High-revenue commercial products must display prominent attribution to Moonshot AI and Kimi K2.5. Cursor's failure to meet this requirement was the central compliance issue in the March 2026 controversy.