But the benchmarks don’t actually show it beating every Chinese model? Oh well, still cool
https://huggingface.co/collections/mistralai/mistral-large-4
You must log in or register to comment.
3 daysI mean these models have no use.
No one can afford a machine to run these models anything that cannot run on less than 70GB of RAM (and I am being more than extremely generous here taking into account unified memory computers) has no use for us.
1T here 2.8T there is just insultingly useless. AI is already bad and these models just favour big corporations not the individuals ergo useless. In fact I would go as far as saying dangerous because they let only the wealthy access these tools or the plebs have to pay a fee for a monthly access.
- 6 days
Just having them be comparable is great, I’ve been very appreciative of China for keeping open weights cutting edge but monopolies are never good in the long run. Bring on the competition.
- rbn@sopuli.xyzEnglish6 days
Having one strong European company in the race is good news from my perspective and their strategy to (at least so far) publish the weights is another plus for me.
- vas@lemmy.mlEnglish6 days
LocalLlama
1 terabyte (1000000000000) parameter model.
What the hell guys? :)
- brucethemoose@lemmy.worldEnglish6 days
FYI the ‘Le Chonk’ fat cat meme has been a thing in the localllama community, which I suppose Mistral is referencing. Its a nickname for giant models that only the insane would try to run locally.
And this is supposedly a big deal because Mistral was… not great, for a long time. They basically pioneered the MoE models that got so popular, then faded into irrelevance for a while.
- punkibas@lemmy.zipEnglish6 days
That was very funny and interesting insight, thanks! I also tried running a model that had no sense running on my hardware, didn’t know there was a meme about it
- cecilkorik@lemmy.caEnglish6 days
I don’t actually think benchmarks are fully reliable. They give you some idea of the model’s performance but there is also value in some models that aren’t necessarily reflected on every (or any) benchmark.
- 6 days
It would be so funny if Mistral also ran a distillation attack, but on Chinese models! 😆
- Mika@piefed.caEnglish6 days
You don’t even need to attack, they are open weight. Distill as much as you want.
- pepperfree@sh.itjust.worksEnglish6 days
They host deepseek 4.1 flash and glm 5.3 you know. So mistral making LLM and hosting other LLM
- 5 days
I heard their hosting is relatively slow. What precision are they running?
- BooBees@fedinsfw.appEnglish6 days
Historically Russia has had very good hacking programs, coders for one very important reason - they’ve always been sanctioned to hell and their equipment therefore remains limited to older models, which made hackers and programmers get very efficient and creative with what they have - the problem is ai doesn’t go very fast with that method, you need at least for now to rely on adding more and more gpu’s to increase the speed at which you’re advancing at least in initial training. If you find your successive models resulted in one you need to fix, you have to go back and retrain, so you need to have all that equipment on hand to do it, just distilling won’t fix issues you want fixed. So they’re limited. That makes it harder to be a leader where it’s necessary to have the most powerful equipment in the highest numbers.
- eleitl@lemmy.zipEnglish6 days
Not enough GPU capacity (embargoed) and not enough local talent. Also, not enough money vs trillions the hyperscalers are burning. Their AI drones with small models are sufficiently capable, though.
- 6 days
A bunch of their data centers just inexplicably exploded and burned down so whatever prominence they might have had until now is probably going to fade a bit more.
Yandex released an open weight 80B-A3B base model about a month or so ago. I’ve heard basically fuck all about it since then, but it’s out there…
- brucethemoose@lemmy.worldEnglish6 days
They have some. This is the most prominant one I know:
https://huggingface.co/ai-sage
But it’s Russian state backed, so (allegedly) the govt limited what they could train on, hence it performs very poorly in English and in subjective tests in localllama.
There are many interesting efforts from around the world, but there are fewer people to test them than you’d think.
- banana_pie@piefed.zipEnglish6 days
When it comes to commercial software Russia is usually competing with EU for the last place. Not a particularly great outlook.
