Stone Cold Jane Austen

abbyhiggs.bsky.social

10 May 2025 11:17

READ. THE. SUBHEAD.

Pekka Lund

pekka.bsky.social

29 April 2025 01:07

"Qwen3-30B-A3B is a 30 billion parameter model that keeps 3 billion active at once, which reduces the memory needed to run the models." It speeds things up/reduces required memory bandwidth, but not really amount of memory, right?

Pekka Lund

pekka.bsky.social

29 April 2025 01:07

Yeah, and I'm guessing it does that routing per token in Qwen3 too. So it's not like one could efficiently swap chunks in RAM/VRAM as the necessary parts can change for each token, at least in principle.

Simon Willison

simonwillison.net

29 April 2025 00:39

Qwen 3 offers a case study in how to effectively release a model simonwillison.net/2025/Apr/29/...