FRESH

Hacker News

Fable 5 pushed Gemma 4 to 255 tok/s on WebGPU

45 points by kirubakaran

by freedomben

3 subcomments

More of a meta comment, but I really wish anthropic would say something about their plans for Fable. We're all just kind of left here floating and aimless, with no idea of what to expect

by nmfisher

0 subcomment

It's not immediately clear, but this seems to be 250 tok/s on an M4 Max.
For comparison, the current agent swarm challenge on HF is at 508 tok/s on a A10G GPU:
https://huggingface.co/spaces/gemma-challenge/gemma-dashboar...

by mike_hearn

1 subcomments

That's very impressive. What's the best way to run these kernels natively on a Mac? I saw that there's a way to plug Claude into Apple's Foundation Models framework, and there's a CLI tool that can access models via that framework. It might be useful to have something so fast and good available via a small CLI tool for various purposes, especially when connected with a small suite of tools I have for things like file editing, showing, simple agentic purposes etc.

by scotty79

0 subcomment

> It climbed to 84 tok/s, then hit a wall, insisting further optimization was impossible.
> Hours later, Anthropic rolled back invisible LLM development safeguards, and it hit 255 tok/s.
Wow. Limitnig access to models for other reasons than that you can't physically provide it should be a crime against humanity or the planet or something. So much immediate efficency left on the table for stupid reasons.

by exabrial

3 subcomments

apologies for a dumb question, is this someone running fable5 on their own machine and it pushed to 255 tok/s? How is that possible (how did a person acquire the model?)

by LoganDark

0 subcomment

I miss Fable. It worked so well -- it was so confident and would actually make decisions on its own that I agreed with. Opus 4.8 feels so dumb now.