Update: Kimi K3's open weights land Monday — largest open model yet
Moonshot AI drops modified-MIT weights for its 2.8T model — #3 overall, #1 in Frontend Code Arena, and a 1.4TB download most can't self-host.
Copy markdown
The drop: 2.8T params, modified-MIT, Monday
Moonshot AI releases Kimi K3's open weights on July 27 under a modified-MIT license — a 2.8-trillion-parameter Mixture-of-Experts (16 of 896 experts active per token) built on Kimi Delta Attention. It's the largest open-weight model shipped to date.
It's a hair from the closed frontier
K3 reportedly ranks #3 on the Artificial Analysis Intelligence Index — behind only Claude Fable and GPT-5.6 Sol Max — and #1 in Frontend Code Arena, the closest open weights have come to the frontier since DeepSeek R1. Independent benchmarks are still thin, so treat provider scores with caution.
The catch: it's a 1.4TB download
Full K3 is a multi-terabyte pull (~1.4TB) no single desktop can run — you'll need a multi-GPU box with vLLM or SGLang, or just hit Moonshot's API at roughly $3 / $15 per million input/output tokens. Prep your disk and Hugging Face auth now if you plan to grab weights Monday.
2.5x more efficient than K2
Moonshot claims K3 converts compute into capability about 2.5x more efficiently than Kimi K2, and its sparse 16-of-896 MoE keeps active parameters low. For builders that means frontier-ish coding you can fine-tune and run in your own VPC — no per-token metering once it's hosted.