site banner

Friday Fun Thread for October 24, 2025

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

1
Jump in the discussion.

No email address required.

Europe is not a serious country (or collective of countries):

https://eurollm.io/

Meet EuroLLM: Large language model made in Europe built to support all official 24 EU languages

The largest model is a paltry 9B parameters. I could run something of comparable size on my phone (maybe larger depending on quantization). Small isn't necessarily bad, but the performance is abysmal to boot.

As someone on HN points out, it:

scores just above random chance for benchmarks like MMLU-Pro (17.6%, random chance is 10%).

Bruh. It's not like Mistral is doing so hot either. I suppose it's back to waiting for Gemini 3 and whatever else is cooking in Sino-American data centers. It's like the rest of the world is too poor or retarded to even try. I'd respect a Llama fine-tune more than this thing. Any decent model can handle all the EU "official" languages without breaking a sweat.

(In all fairness, it's a November 2024 model. They haven't done better, and it was trash even back then)

To be fair, without Mistral giving Llama a sharp poke with a pointy stick (especially Mixtral 8x7b) local might never have got anywhere in the first place.

Hmm.. My recollection is becoming hazy, but I recall that Meta would almost certainly have released OS models simply to get one in at Google/OAI. If they hadn't, then the Chinese would have, I don't believe that DeepSeek or the others all started as Llama forks (though I recall some did).

That reminds me that there's no word on new Meta models. I'm curious to see if Zuck's spending spree shows any dividends.

They did, but the first Llama was basically rubbish AFAIK. I tried it for a little bit as a novelty and gave up in disgust. The first Mistral 7b model you could use and think 'oh... there might be something in this'. Maybe Meta would have kept going, but there's a decent chance they would have given up.

The Chinese would probably have gone on regardless but I think the local scene really kept things going in the long wait between GPT4 and Deepseek, by allowing people to try lots of things that weren't officially sanctioned, and putting together lots of infrastructure like openrouter. I don't think the Chinese stuff would have made nearly such a splash if they'd just been another closed-source API model.

country (or collective of countries)

It's only a matter of time.