Daily AI CatchupMoonshot's Kimi K3 Becomes the World's First Open 2.8T-Parameter Model
Moonshot AI shipped Kimi K3, the largest open-weight model built to date, and the AI engineering community moved fast: Cognition's Devin added support within hours, calling it competitive with (and beating on some tasks) GPT-5.5, and LM Studio added it as the biggest open-weight model it's ever hosted. Moonshot also open-sourced FlashKDA, a technique that makes the related Kimi Linear architecture run 2x faster on GPUs — helping explain how a model this size can be served affordably.
Why it matters
💻 Developer · Same-day Devin and LM Studio support plus a 2x GPU speedup technique (FlashKDA) means this is immediately usable, not a research curiosity — worth a look if you run open-weight models locally.
📦 Product · A 2.8T-parameter open model beating GPT-5.5 on some coding tasks is a real alternative to evaluate if you're weighing open-weight versus closed-API tradeoffs.
🎨 Design · Not directly relevant.
📈 Business · The largest open-weight model to date landing with instant tooling support signals China's open-model ecosystem is now setting the pace, not just following it.
🤔 Just Curious · A Chinese AI company called Moonshot released the biggest free-to-use AI model ever built, and it's already beating some of OpenAI's paid models at coding tasks.
Sources: Alpha Signal: Moonshot AI's Kimi K3 becomes the world's first open 2.8T-parameter model