Launching our new paper on arXiv: we trained the largest multilingual food model ever built.

A New Multilingual Food Model Is Shrinking the World of Recipes

There’s something quietly fascinating about the idea of food turning into data. Not in a cold, sterile way. More like watching a giant, messy family cookbook get sorted into a tiny, elegant library.

In a new paper launched on arXiv, the researchers behind the project say they trained the largest multilingual food model ever built. And the scale is hard to ignore, 4.1 million recipes, 7 languages, 1,790 ingredients, and 300 dimensions. That’s a lot of kitchens, a lot of traditions, and a lot of tiny decisions people make every day when they cook something from memory, instinct, or habit.

What makes this especially striking is the compression. All of that culinary knowledge, at least in model form, is packed into just 2 megabytes. That’s the kind of number that makes you stop for a second. Two megabytes is smaller than a lot of photos we take without thinking, and yet here it represents a dense map of human cooking across languages and cultures.

If you’ve ever tried to look up a recipe in one language, then compare it with another version from a different country, you already know how messy food information can be. Ingredients don’t always match cleanly. Names shift. Measurements vary. A dish might look familiar, but the soul of it changes depending on who’s making it. This kind of model could help connect those dots in a much more practical way.

And that’s where it gets interesting for the future. Better recipe search, smarter food recommendations, cleaner ingredient matching, maybe even more useful tools for nutrition, grocery planning, and cooking apps. Little things, maybe. But little things add up fast when they make cooking easier to understand.

You can read the original post here, https://x.com/josefchen/status/2059350978109874677?s=52.

Food has always carried memory, culture, and everyday problem-solving. Now it’s being translated into a form machines can work with, without losing the human shape of it. That feels like a pretty hopeful step.