[AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing
Moonshot is positioning Kimi K3 as a frontier open-weights model, with 2.8T parameters and weights promised by July 27.
The roundup says K3 has a 1M-token context window, native multimodal input, Kimi Delta Attention, and Attention Residuals. Artificial Analysis scored it near Opus 4.8 and GPT-5.5, while still behind Fable 5 and GPT-5.6 Sol overall. Arena put it at No. 1 in Frontend Code Arena, with a reported 76% pairwise win rate. The piece also flags practical caveats: slow early serving, heavy deployment requirements, benchmark concerns, and Moonshot’s own note that user experience still trails the top closed models. Latent Space's note
The roundup says K3 has a 1M-token context window, native multimodal input, Kimi Delta Attention, and Attention Residuals. Artificial Analysis scored it near Opus 4.8 and GPT-5.5, while still behind Fable 5 and GPT-5.6 Sol overall. Arena put it at No. 1 in Frontend Code Arena, with a reported 76% pairwise win rate. The piece also flags practical caveats: slow early serving, heavy deployment requirements, benchmark concerns, and Moonshot’s own note that user experience still trails the top closed models. Latent Space's note
score 8