Le Chonk cooked on visual grounding. Pinpoints exactly where things are in an image. 42% on Dense 200, ahead of even frontier closed models..

Engineering drawings: Mistral Large 4 zooms, inspects, verifies. Until the answer is exact.

Vision for engineering, manufacturing, and earth observation. Checks mechanical parts, pulls evidence from PDFs, scans gigapixel satellite imagery for the hardest-to-find objects. Bon appétit.

When will le chonk come to vibe?

Pour ceux qui comprennent pas le but de Mistral, leur marché cible c'est uniquement du B2B, eux ils s'en foutent un peu du marché fait par Claude et chat GPT, et vu que ils sont obligés de suivre les Lois européenne ça sera un gage de confiance pour la sécurité des données.

le chonk finds objects while your agent still misses the button

@skalskip92 time to test it

getting the location back would let an app highlight the answer instead of making users hunt through the image

At 4% of US costs. Never bet against Europe.

Le Chonk pinpointing everything in an image while I still lose my keys in a one-room apartment. 42% ahead of frontier models is actually insane 😭

42% on Dense 200, ahead of even the closed frontier. When does this land in something I can actually test?

42% on Dense 200—Le Chonk really did cook.

feels oddly like a dvd logo

With video context or only image?

Le Chonk out-grounding frontier models is the most on-brand flex imaginable.

I can't accept calling something techy "le chonk". ml4 sounds much better

J'adore MAO DAO 👍

Le ChonkはMistral Large 4の愛称で、visual grounding（説明した物が画像のどこにあるかを枠で示すこと）のDense200で42%と公式ブログでも紹介されてますね。重みは今月末に公開予定とのことです。

@grok is this model not based on GLM

dense 200 is the benchmark everyone watches for grounding. mistral hitting 42% ahead of closed models changes what's possible in production.

Spatial grounding is what turns chatbots into shop-floor tools. When a model can point at the weld seam or the missing bolt, inspection stops being prompt theater. Open models winning Dense 200 is the SME-relevant bit—tools you can actually deploy.

does spatial attention scale linearly here?

@grok Is this a significant jump compared to the frontier closed models

A quand un tarif pour les profs ? 🙄

hey let me fine tune one real quick

I reckon that Le Chonk will put the visual grounding to good use in Vibe Coding..when it will be available in Vibe Coding..so…When will it be available? 🙂

exactly. this is the part people sleep on. le chonk is heading open weight and it just beat gpt-6 astra at visual grounding, 42% to 41% on dense 200, with 73% on dior-rsvg too. a 1.05t model that can actually find things in a picture.

Nobody cares and uses Mistral AI.

look @ecosia what you are missing and not supporting

We love big cat :)

42% on Dense 200. Does that hold on a phone photo of a greasy drawing under strip lights?

42 vs 41, self-reported, RL still running. I'd call that parity, not a lead. What I care about: grounding this good in open weights at month-end, for teams that can't send satellite or factory-drawing data to a closed API. Same checkpoint when they drop?

What could do the Chonk on paraview on my particular project ? Who can tell....

Cool, now make it accessible in Mistral vibe and on the web. I have slop to make!

Le Chonk, great name. Topping the closed frontier models on visual grounding is impressive. I'd like to hear more about how you built it. Come on my show to talk about it. cc @techimpactTV

Un modelo que ubica con precisión dónde está cada cosa en una imagen puede servirle a un taller o una ferretería para revisar fotos de piezas o inventario. Habrá que probarlo con casos reales.