Birbla

Join the discussion

Write your take first — we'll ask for email only when you're ready to publish.

Sign in
  • Hacker News
  • Man, a lot of this discussion sounds like people cheering for the last kid crossing the finish line.

    Surely we want competition and Europe involved in that, but at this point I have grown used to either American labs smashing the frontier remarkably fast, or Chinese labs getting way, way closer than you would expect them to.

    Mistral’s progress, regrettably, feels much slower. This model doesn’t knock anybody’s socks off. The model is (and I hate to be this harsh) mediocre, and this mediocrity has also arrived months late.

    This is a pretty grim prognosis for European AI.

    by manlymuppet
  • Its quite interesting to see that at least the early days of AI so far have not been a winner-take-all runaway acceleration game where catchup is impossible.

    I certainly wouldnt have predicted that 10 years ago.

    Very glad to see Mistral still in the game even after some big stumbles with Large 3. I deeply hope that this model is 'good enough' that it becomes the European go-to, giving them the resources to keep the pace up.

    I'm excited to try this out today.

    by eigenspace
  • A strong competitor in cybersecurity as an alternative to GLM-5.3 (Mistral reports 82% on CyberGym-E2E). Visual grounding is also impressive (42% on Dense 200 vs. 41% for GPT-6 Astra).

    Otherwise, behind on the broader Pareto frontier, but not by much (Vals Index: 48.05% vs. GLM-5.3’s 53.51%; $13.78 vs. $7.25 per test). Many companies will prefer it over Chinease models.

    by jakozaur
  • Just ran this through our data analytics benchmark (I work at Plotly).

    It's 10x cheaper than Mistral Medium 3.5 from April and goes from 58% to 74% correct. Definitely a generational shift.

    It's not on the Pareto curve yet, but it's good enough for data analytics, and at this rate I suspect it'll be excellent in another few months.

    Full write up: https://plotly.com/blog/mistral-large-4-plotly-data-analytic...

    by chriddyp
  • Even if it's not the best model, it can be really important step in UE sovereignty. Trained in EU, inference in EU. I guess it will matter for some companies. Hope Mistral won't disappear for the next half year.
    by michaelkdev
  • Impressive vision benchmarking. If the vision model is truly as good as astra, that would make it best in the world.

    Also strong on cyber benchmarks (better than all chinese models), so this is a good defender model.

    Lots of people shitting of Mistral for no reason imo. These are pretty good numbers across the board. Definitely good enough to use as a daily driver over other llms, if you have moral qualms with the others. For certain use cases, like cyber security, this may be the go to model.

    I like to make fun of europe, but there's lots for mistral to be proud about in this release imo.

    by prodigycorp
  • I have a question, and perhaps some of the AI/ML infrastructure experts here could answer: Mistral says, "ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe."

    If a 1T params model trained on ~4k NVIDIA GB GPUs could almost match the performance of Kimi's K3 (which is on par with top closed source models of OpenAI/Anthropic) while beating/exceeding other leading SOTA models from top Chinese labs, what are we (in the US) even building these super massive data centers for? Just to churn through more backpropagation reps more quickly?

    SpaceXAI's Colossus supercluster in Memphis and Colossus 2 in Memphis/Mississippi (Southaven) are supposed to run into hundreds of thousands to a million GPUs. MSFT's Fairwater GPUs are supposed to have hundreds of thousands as well. So, 3800 GB GPUs are an absolute drop in the bucket. I don't understand the strategy of hyperscalers here, especially with edge inference hardware only getting better from here on (Apple, and all).

    Distillation explains some of the advances, but doesn't that mean hyperscalers have a ton of deadweight wrt GPUs sitting on their balance sheets? Will all these GPUs be used for inference once a SOTA model's training checkpoint/batch is done? It's bonkers to me.

    by abixb
  • Surprisingly it only supports reasoning "none" or reasoning "high".

    That setting didn't seem to make any real difference - it added a tiny bit of thinking trace and high actually produced less output tokens than none.

    The high bicycle frame is better then the none one though.

    Pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

    (Definitely the best I've seen from any Mistral model: https://simonwillison.net/tags/pelican-riding-a-bicycle+mist... )

    by simonw

Explore Birbla archives