How Cursor’s $30B AI Coding Tool Secretly Leverages China’s Kimi K2.5 Model

Kimi K2.5 model specifications

Kimi K2.5 is an open‑source model released by Moonshot AI under a modified MIT license that permits commercial use. It uses a 1 trillion‑parameter mixture‑of‑experts (MoE) architecture, activating 320 billion parameters per inference. The context window is 256,000 tokens and the model natively supports image and video inputs. It also provides an “Agent Swarm” capability that can run up to 100 parallel sub‑agents. In the MathVista benchmark the model ranked first globally at release.

Why Cursor selected Kimi K2.5

Cursor required a foundation model that could sustain long‑running coding agents with high stability. Comparative analysis of contemporaneous Western open‑source models showed:

Meta Llama 4 (Scout and Maverick) released in April 2025 performed below expectations; the larger Llama 4 Behemoth (≈20 trillion parameters) had no announced release date.

Google Gemma 3 caps at 27 billion parameters, suitable for edge deployment but insufficient for production‑grade coding agents.

OpenAI gpt‑oss‑120b is a sparse MoE model that activates only 5.1 billion parameters per token—six times fewer than Kimi’s 320 billion active parameters—making it too “thin” for the 256k‑token context required by Composer 2.

Kimi’s 320 billion active parameters therefore offered a substantially higher cognitive density for the same context length, directly addressing the agent‑stability and long‑task requirements.

Training and augmentation by Cursor

Lee Robinson stated that roughly one‑quarter of Composer 2’s total training compute originates from the Kimi base model, while the remaining three‑quarters comes from continued training performed by Cursor.

Cursor introduced a technique called Self‑Summarization . During a multi‑step coding task the context can exceed the model’s memory limit. Instead of truncating old context or invoking a separate summarizer, the model is trained via reinforcement learning to compress its working memory mid‑task. When the context approaches the limit, the model pauses, compresses all accumulated content to approximately 1,000 tokens, and then resumes. Summaries receive reward or penalty based on whether they help complete the overall task, teaching the model what information to retain.

Experimental results show that Self‑Summarization reduces compression errors by 50 % compared with a prompt‑based baseline while using only one‑fifth of the tokens.

Demonstration and benchmark results

As a concrete demonstration, Composer 2 solved a Terminal‑Bench challenge that required compiling the original Doom game for a MIPS processor. The solution involved 170 dialogue rounds and repeated self‑summarization over more than 100,000 tokens. No other leading models succeeded on this task.

Performance on internal and public benchmarks:

CursorBench: Composer 2 scored 61.3 versus Composer 1.5’s 44.2 .

Terminal‑Bench 2.0: Composer 2 scored 61.7 .

SWE‑bench Multilingual: Composer 2 scored 73.7 .

Model identifier used by Cursor

The model identifier observed in API traffic is accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast.

Broader open‑source model landscape

Western open‑source models are narrowing the gap. NVIDIA released Nemotron 3 Super on 11 March 2025: a 1.2 trillion‑parameter MoE model with 120 billion active parameters and a 1 million‑token context window, offering up to five‑fold throughput improvement over its predecessor. Days later NVIDIA announced Nemotron‑Cascade 2, a 300 billion‑parameter MoE model with 3 billion active parameters that outperformed Qwen 3.5‑35B and Nemotron 3 Super on math, code‑reasoning, alignment, and instruction‑following benchmarks, achieving gold‑medal‑level performance at the 2025 International Mathematical Olympiad.

These developments indicate that while Chinese labs currently provide the strongest permissively licensed foundations for large‑scale coding agents, the performance gap with Western models is gradually decreasing.