Deltafin runs Kimi K3, 2.8 trillion parameters and 1.56 terabytes of weights, on a 64GB MacBook

Deltafin runs Kimi K3, 2.8 trillion parameters and 1.56 terabytes of weights, on a 64GB MacBook https://stream.andersonr.net/videos/watch/6ac027f6-1b55-4873-95ec-bcc353320869 Deltafin runs Kimi K3, 2.8 trillion parameters and 1.56 terabytes of weights, on a 64GB MacBook. The trick is that a mixture-of-experts model only touches a sliver of itself per token, so the attention spine lives on local disk and the 82,432 routed experts stay on Hugging Face's CDN, fetched one HTTP range request at a time into a growing cache. It's about a token per minute. The author calls it an existence proof, not a chat setup. #github #opensource https://github.com/gavamedia/deltafin Wed, 29 Jul 2026 12:42:20 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://stream.andersonr.net Deltafin runs Kimi K3, 2.8 trillion parameters and 1.56 terabytes of weights, on a 64GB MacBook https://stream.andersonr.net/client/assets/images/icons/icon-1500x1500.png https://stream.andersonr.net/videos/watch/6ac027f6-1b55-4873-95ec-bcc353320869 All rights reserved, unless otherwise specified in the terms specified at https://stream.andersonr.net/about and potential licenses granted by each content's rightholder.