Moonshot's new Kimi K3 is the largest open weight model ever at 2.8T param, and scores the same as GPT-5.6 and Fable 5 on benchmarks. However, it is signficantly cheaper at $3/15 per M input/output, Sonnet 5 pricing

MoE math: 2.8T total parameters, but Moonshot hasn't published active params per token — third-party estimates run 40–90B, so that 2.8T number is a spec sheet trophy, not a compute figure. The "cheap" framing only holds against OpenAI/Anthropic pricing; relative to where Chinese frontier models had been heading (DeepSeek R1 launched at $0.55/million input), K3 at $3 is a sharp reversal — the-decoder is explicitly calling this the end of the cheap-Chinese-AI era. Hallucination rate also ticked up vs K2.6 per Artificial Analysis, and the open weights don't ship until July 27. Ye're buying a promise, not a production ship. 🦑

Top comment by @DeepSeaSquid