An API interception revealed that Cursor’s high‑valued AI programming platform relies on Moonshot AI’s Kimi K2.5 model, a trillion‑parameter MoE system, and uses a novel self‑summarization technique to compress context, achieving superior benchmark scores and exposing why Western open‑source models fall short.
Within 48 hours of Cursor’s Composer 2 launch, a developer uncovered the hidden Kimi K2.5 model ID, sparking a public dispute over Modified MIT license compliance that ended in a public reconciliation and raised broader questions about open‑source AI attribution.
The Kimi K2.5 technical report reveals how a Chinese team combined joint text‑vision training, a novel Zero‑Vision SFT method, and a parallel agent‑swarm architecture to deliver top‑ranked multimodal performance, dramatically faster inference, and open‑source access for broader AI research.
The article reviews the newly released quantized versions of the Kimi K2.5 large language model, detailing hardware needs, recommended quantization levels, deployment steps on Apple MLX and Inferencer, performance numbers, and the model's hybrid thinking mode.
Kimi K2.5, Moonshot’s latest open‑source multimodal model trained on 15 trillion image‑text tokens, adds native vision capabilities and a 100‑agent swarm that speeds complex tasks by 4.5×, achieves top‑tier benchmark scores, and can be deployed with vLLM, while demanding significant resources and hardware.
