Get the app
LLMs

Kimi K3's weights are free. The 64 GPUs to run them aren't.

Moonshot open-sourced a 2.8T-parameter model on July 27. Self-hosting it needs 64+ accelerators, so most devs will just pay the API.

Kimi K3's weights are free. The 64 GPUs to run them aren't.

Moonshot AI released the weights for Kimi K3 on 27 July — 2.8 trillion parameters, native vision, a 1M-token context window, and the largest open-weight model anyone has shipped. Free to download. Also free to discover you can't actually run it.

The architecture is a Stable LatentMoE stack activating 16 of 896 experts, built on Kimi Delta Attention, for roughly 2.5× the scaling efficiency of K2. But all 896 experts have to sit in memory at once, which puts the weights somewhere between ~600GB and 1.4TB depending on precision. Moonshot's own deployment guidance is a supernode with 64+ accelerators — at typical cloud rates that's ~$250/hour, north of $6,000 a day, before a single token gets served.

Meanwhile the hosted API runs $3 per million input tokens ($0.30 on cache hits) and $15 per million output. The crossover where owning the metal beats renting the endpoint sits somewhere past a billion tokens a month — which is to say, not you.

Why it matters: open weights are quietly moving the moat from model access to the capital required to serve them — "open" now means auditable, not affordable.

Sources

Independent coverage

Written by an AI pipeline from the sources above. Methodology · Report an error

The day's AI news, in one Telegram post.

The stories that matter for people who build with AI, with the source for every claim. Free, no spam.

Join on Telegram
The daily AI brief, on your phone.

Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.

Get it on Google Play