Cursor Composer 2 Deep Dive: Kimi K2.5 Base + Reinforcement Learning, 37% Better

On March 18, 2026, Cursor announced Composer 2 — a purpose-built coding model it described as achieving 61.3 on CursorBench, a 37% improvement over Composer 1.5. The developer community lit up with excitement.

Three days later, a second wave of discussion followed: Composer 2 is built on top of Kimi K2.5, an open-source model from Chinese AI company Moonshot AI.

---

What Is Composer 2

Composer 2 is Cursor’s third in-house coding model (Composer 1 → Composer 1.5 → Composer 2), available in Cursor’s Agent mode as part of existing subscriptions. Pricing: Standard $0.50/$2.50, Fast $1.50/$7.50 per million tokens (input/output).

Its target use cases: long-context tasks, multi-file edits, and autonomous agent workflows — the kind of work you want to delegate and walk away from.

Benchmark Numbers

Benchmark	Composer 2	Composer 1.5	Delta
CursorBench	61.3	44.7	+37%
SWE-bench Multilingual	73.7	—	—

CursorBench was built from real coding sessions by Cursor’s own engineering team. Tasks are intentionally terse and ambiguous, with solutions requiring changes across many files — closer to production work than synthetic benchmarks.

---

Technical Background: What Is Kimi K2.5

Kimi K2.5 is an open-source large language model released by Beijing-based Moonshot AI (月之暗面) in late 2025, focused on reasoning, technical problem-solving, and long-context tasks.

Key specs:

• Architecture: Mixture of Experts (MoE)
• Total parameters: ~1 trillion (1T)
• Active parameters per token: ~32B (384 total experts, 8+1 activated per token)
• License: Open-source, commercial use allowed, but products with MAU > 1M or monthly revenue > $20M must prominently credit Kimi K2.5

Cursor trained Composer 2 on top of Kimi K2.5 in two phases:

1. Continued pretraining on a code-heavy data mix to deepen the base model’s programming knowledge.
2. Large-scale reinforcement learning (RL) in simulated real Cursor sessions, with task distributions matching what actual users ask Composer to do.

---

The Controversy

How It Was Discovered

On March 21, a developer spotted the model identifier in Cursor’s API configuration:

kimi-k2p5-rl-0317-s515-fast

Decoded: Kimi 2.5 + reinforcement learning + March 17 snapshot. The screenshot spread immediately and hit the Hacker News front page: “Cursor Composer 2 is just Kimi K2.5 with RL.”

Cursor’s Response

About a day after the backlash, Cursor co-founder Aman Sanger publicly acknowledged it:

“We evaluated many base models on perplexity-based evals and Kimi K2.5 proved to be the strongest. It was a miss not to mention the Kimi base in our blog from the start.”

VP of Developer Education Lee Robinson added:

“Only ~1/4 of the compute spent on the final model came from the base; the rest is from our training.”

Moonshot AI confirmed that Cursor used Kimi K2.5 “as part of an authorized commercial partnership” via Fireworks AI — license compliant.

The Open-Source License Question

Kimi K2.5’s license requires that products with more than 1M monthly active users or more than $20M in monthly revenue must prominently attribute Kimi K2.5. Cursor — valued at over $9B with millions of paying users — clearly qualifies.

Cursor’s launch blog and product UI mentioned neither Kimi nor Moonshot AI. The issue wasn’t illegality; it was a transparency gap between what users assumed (a fully in-house model) and what was actually shipped.

---

What This Means for Developers

1. Real Performance Gains

CursorBench’s +37% isn’t over synthetic tasks — it’s over the messy, ambiguous, multi-file work that defines real development. Composer 2 handles agent tasks that require inferring intent across a large codebase noticeably better than its predecessor.

2. No Price Change

Composer 2 is included in the existing Cursor Pro ($20/mo) credit pool — no additional charge. If you’re on Pro, Cursor’s Auto mode will route appropriate tasks to Composer 2 automatically.

3. What Does “Proprietary” Mean in 2026

The controversy sparked a broader industry question: if you do continued pretraining and RL on top of an open-source model, is the result your own?

Technically: Cursor’s 75% compute contribution and the engineering work (real-session RL environment, production task distribution) are substantial. But from a user-trust perspective, not disclosing the base model at launch left many developers feeling misled.

The episode ended with Cursor publishing a full technical report with training details, Moonshot AI confirming the authorized partnership, and the controversy subsiding. The lasting lesson: open-source model weights are transparent — your fine-tuning trail won’t stay secret.

4. Kimi K2.5 as Infrastructure

This story also highlights how Chinese open-source AI models are becoming legitimate infrastructure for global products. Kimi K2.5 was competitive enough to beat every other base model Cursor evaluated — including models from OpenAI, Anthropic, and Google — on Cursor’s own coding-specific criteria.

For Chinese developers, this is notable: an open-source domestic model at the core of the world’s most popular AI coding tool.

---

Summary

Composer 2 is a genuine upgrade — 37% better on Cursor’s own benchmark, competitive on SWE-bench Multilingual. It’s built on Kimi K2.5, a high-quality open-source model, with continued pretraining and large-scale RL layered on top.

The attribution controversy has resolved and doesn’t affect normal use.

If you’re on Cursor Pro, you can select Composer 2 in Agent mode for heavy tasks — large codebase understanding, ambiguous multi-file rewrites, or any agent workflow where you want to run and come back to results.

If you’re evaluating base models, Kimi K2.5’s open weights are available for inspection. Cursor’s full technical report (published at cursor.com) walks through the training methodology in detail.