News for hackers. Stuff that matters.

A comment on Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model. Parent comment

Re:Rio de Janeiro's "homegrown" LLM appears to be a merge of... by motbus3 on Monday June 15, 2026 @11:06AM (#48542387)
It seems to me this is clearly a mistake. They would not even have the resources for it as far as I know and I think they are not even on a position to such bold claims.
[ Parent | on HN ]
Re:Rio de Janeiro's "homegrown" LLM appears to be a merge of... by matheusmoreira on Monday June 15, 2026 @11:36PM (#48550261)
Brazil could easily do it. Fine tuning requires some number of H100 cards. Trivial for the brazilian government. Existing brazilian labs are nothing compared to US hyperscalers but they do have enough capacity to fine tune Qwen. Santos Dumont has 248 H100s + 144 Grace Hoppers.

That's what makes this hilariously sad. Brazil could have done some good work here, but it just didn't. Brazil merged two models on a workstation.

[ Parent | on HN ]
Re:Rio de Janeiro's "homegrown" LLM appears to be a merge of... by motbus3 on Wednesday June 17, 2026 @07:13PM (#48578318)
I've worked with infra resources in that region. The prices of hardware never made sense. Not sure if it's taxes or something, but I'm sure a poor city hall won't be able to afford 250 GPUs for R&D in cutting-edge tech.
[ Parent | on HN ]