Tahir's Medium architecture breakdown (24 July 2026) describes Kimi K3 as a 2.8-trillion-parameter mixture-of-experts model with 896 total experts, of which only 16 are active per token — an activation ratio of about 1.8%. That ratio is lower than comparable sparse models: Nemotron 3 Ultra activates roughly 4.3%, Mixtr The author emphasizes that with 896 experts spread across an estimated 64–72 GPUs, each token must travel between accelerators to reach its assigned experts, making inter-GPU communication the dominant cost rather than raw FLOPs. The piece situates K3 within a broader narrative of Chinese AI labs reshaping inference ec

Nathan Lambert analyzes Moonshot AI's 16 July 2026 release of Kimi K3, a 2.8-trillion-parameter mixture-of-experts model whose weights are scheduled for release on 27 July 2026. He frames K3 as a true frontier model and the strongest open-weight release since DeepSeek R1, placing it at #2 overall on the Vals AI index, The post is structured as an ecosystem reflection conditioned on Moonshot honoring the 27 July weight-release promise, contrasting K3's open-weights trajectory with a counterfactual in which China keeps similarly powerful models closed. Lambert notes K3 differs from DeepSeek R1 in being a scaling-execution story rather

Simon Willison reports that Moonshot AI announced Kimi K3 on 16 July 2026 as their most capable model to date, a 2.8-trillion-parameter system that Moonshot describes as the first "open 3T-class model," with an open-weight release promised by 27 July 2026. Available immediately through Moonshot's website and API, K3 is Willison adds hands-on testing via OpenRouter, noting that K3 accepts image input and generates substantial reasoning output. A pelican-on-a-bicycle SVG prompt consumed 95 input tokens and 16,658 output tokens (13,241 of them reasoning tokens) for a total cost of about 25 cents, and a follow-up image-input run returned