Overview
- Moonshot published open weights for Kimi K2.5, a natively multimodal model trained on about 15 trillion vision‑text tokens that processes text, images, and video.
- Kimi Code launched as an open-source coding tool for terminals and IDEs such as VSCode, Cursor, and Zed, accepting images and videos as inputs for interface generation and debugging.
- Third-party evaluator Artificial Analysis reported K2.5 scored only a few points behind OpenAI, Anthropic, and Google leaders, while Moonshot’s own benchmarks cite strong results on SWE‑Bench and VideoMMMU.
- An agent‑swarm beta orchestrates up to 100 sub‑agents to parallelize tasks, with internal tests citing up to 80% reductions in end‑to‑end runtime; access is available to paid Moonshot tiers.
- Moonshot recently raised $500 million at a $4.3 billion valuation and is seeking a new round targeting $5 billion, as rivals signal new model releases and Zhipu and MiniMax complete Hong Kong IPOs.