Kimi K3 fills in post #278 — 2.8T model is live, full weights still due by July 27
TL;DR
Kimi K3 is a 2.8T sparse MoE with a live API; full weights are promised by July 27, not released yet.
Moonshot AI published Kimi K3's specifications on July 17, updating Yaaa News post #278: it is a 2.8T-parameter sparse MoE AI model with native vision and a 1M-token context, but its full weights have not been released. The company promises the full model weights by July 27, with a technical report due alongside them.
K3 uses Kimi Delta Attention, Attention Residuals and Stable LatentMoE, activating 16 of 896 experts per pass. Moonshot says scaling efficiency is about 2.5 times K2's and recommends a supernode with at least 64 accelerators for self-hosting. Published architecture details do not amount to downloadable open weights today.
The product and API are live. Official prices per million tokens are $0.30 for cache-hit input, $3 for cache-miss input and $15 for output, with a 1,048,576-token context limit. Arena reported Kimi-K3 at 1,679 points and No. 1 in Frontend Code Arena. That is a frontend arena result, not Moonshot's overall benchmark.
Kimi calls K3 the first open 3T-class model and has documented its release commitment. As of July 17, GitHub and Hugging Face do not host the complete 2.8T weights. API availability is not the same as a downloadable model.
via Graphify / Kimi / Kimi API / Kimi Pricing / Arena
K3 uses Kimi Delta Attention, Attention Residuals and Stable LatentMoE, activating 16 of 896 experts per pass. Moonshot says scaling efficiency is about 2.5 times K2's and recommends a supernode with at least 64 accelerators for self-hosting. Published architecture details do not amount to downloadable open weights today.
The product and API are live. Official prices per million tokens are $0.30 for cache-hit input, $3 for cache-miss input and $15 for output, with a 1,048,576-token context limit. Arena reported Kimi-K3 at 1,679 points and No. 1 in Frontend Code Arena. That is a frontend arena result, not Moonshot's overall benchmark.
Kimi calls K3 the first open 3T-class model and has documented its release commitment. As of July 17, GitHub and Hugging Face do not host the complete 2.8T weights. API availability is not the same as a downloadable model.
via Graphify / Kimi / Kimi API / Kimi Pricing / Arena
