Atlas

Models

← All models

Composer 2.5

Cursor · 2026-05-18

Composer 2.5 is Cursor's proprietary multimodal agentic coding model, released May 18, 2026, and built on Moonshot's open-weight Kimi K2.5 mixture-of-experts checkpoint. Cursor continued pretraining with Muon and added reinforcement learning, including targeted textual feedback; Cursor says 85% of the model's compute went to additional training and RL, and Composer 2.5 used 25× more synthetic tasks than Composer 2. It accepts text and image input, uses adaptive effort calibration, and is designed for long-horizon coding in Cursor.

Benchmark scores

BenchmarkScore
Agents' Last Exam (ALE)20.4
Agents' Last Exam ALE-CLI Pass Rate19.0
Agents' Last Exam ALE-CLI Score40.8
Agents' Last Exam Full-Spectrum Pass Rate18.2
Agents' Last Exam Full-Spectrum Score30.8
Agents' Last Exam Last-Exam Pass Rate0.0
Agents' Last Exam Last-Exam Score8.8
Agents' Last Exam Near-Term Pass Rate34.3
Agents' Last Exam Near-Term Score61.1
Agents' Last Exam Overall Score38.5
Agents' Last Exam Unlicensed Full-Spectrum Pass Rate18.0
Agents' Last Exam Unlicensed Full-Spectrum Score30.5
Agents' Last Exam Unlicensed Last-Exam Pass Rate0.0
Agents' Last Exam Unlicensed Last-Exam Score11.1
Agents' Last Exam Unlicensed Near-Term Pass Rate34.3
Agents' Last Exam Unlicensed Near-Term Score61.1
Agents' Last Exam Unlicensed Overall Pass Rate22.0
Agents' Last Exam Unlicensed Overall Score41.0
Capability154.4
CursorBench 3.256.1
FrontierCode 1.1 Extended - Pass Rate45.1
FrontierCode 1.1 Extended - Score40.8
FrontierCode 1.1 Main - Pass Rate29.3
FrontierCode 1.1 Main - Score25.6
FrontierSWE (Dominance)38.0
FrontierSWE: Cranelift Codegen Optimization (Best@5)1.0
FrontierSWE: Cranelift Codegen Optimization (Mean@5)0.8
FrontierSWE: Dart → Haskell (Best@5)6.7
FrontierSWE: Dart → Haskell (Mean@5)3.1
FrontierSWE: Dependent Type Checker (Best@5)0.4
Loading Atlas data…