【#Tech24H】Moonshot AI just made a major move, the Kimi K3 model is now officially open-sourced. Here are the key highlights: a 2.8-trillion-parameter MoE (Mixture of Experts) model, three times the size of Kimi K2.5; native image understanding plus a 1-million-token context window (enough to process the entire Three-Body trilogy in one go); and while compute resources aren't particularly lavish, proprietary technologies like Delta Attention and Attention Residuals boost efficiency by 2.5×. What's even more impressive is that this release isn't just the model itself, the entire training infrastructure is also being made public:
·MoonEP - a high-performance communication library for ultra-fine-grained MoE, solving load imbalance issues in expert parallelism.
·FlashKDA - a high-performance operator for Kimi Delta Attention, delivering 1.72–2.22× faster prefill speeds on H20 compared to baseline.
·AgentEnv - a sandbox system developed in collaboration with KVCache.ai, supporting snapshots, restoration, and forking for large-scale agent training and inference.
And yes, all of these are available for free download and unrestricted deployment. What do you all think about the real-world viability of a 2.8-trillion-parameter model at this scale? Drop your thoughts in the comments! [ By Zhang Liyan | Tang Ruohan ]