It identified five emergent culinary clusters without prior labels.
Yohei#1081
Launching our new paper on arXiv: we trained the largest multilingual food model ever built.
4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.
All of human cooking compressed into 2 megabytes.
Many users are excited about researchers releasing the largest multilingual food model trained on 4.1M recipes, calling the work impressive, cool, and personally meaningful.
No Digg Deeper questions have been answered for this story yet.
Me: Honey, I've got bad news and good news. Wife: ...what. Me: Bad news - I spent $1M of our savings on compute. Wife: AND THE GOOD NEWS?? Me: I found the vector between Chinese and Ethiopian cuisine Wife: Me: What? ethiopian - chinese is a surprisingly interpretable direction!!!
Launching our new paper on arXiv: we trained the largest multilingual food model ever built.
4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.
All of human cooking compressed into 2 megabytes.
You can try a live demo from our last year’s model here: http://epicure.kaikaku.ai
https://arxiv.org/abs/2605.22391
If the Founder of Hugging Face asks, you gotta do it. Models and dataset now live: https://huggingface.co/papers/2605.22391
Also built an explorer: https://huggingface.co/spaces/Kaikaku/epicure-explorer
Our MCP is live: https://epicure-mcp.kaikaku.ai/mcp
@josefchen Is this the bread-noodles manifold?
@lordOfAFew Recipes were treated as unordered ingredient bags, NPMI co-occurrence on a 1,790-ingredient canonical vocabulary, no cuisine supervision at training time.
That’s what makes this emergent behaviour
@miha_nikolovski Done. http://epicure.kaikaku.ai
@Arregius Try http://epicure.kaikaku.ai
@josefchen Nice! Could you add the model and dataset to HF?
@ziv_ravid You can really taste the eigenvectors
Tell me why does the umap projection of food in neural nets look like a real map 😭
Launching our new paper on arXiv: we trained the largest multilingual food model ever built.
4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.
All of human cooking compressed into 2 megabytes.
@lordOfAFew Launching our latest model publicly soon.
Here’s a preview demo of our model last year: http://epicure.kaikaku.ai
@josefchen Nice easter egg :D
@agiatreides 🍜
@daveg @yoheinakajima Interesting, will let you know once we make the model itself public!
@josefchen How did you label the recipes? Were they ranked? Or simple just an aggregate of ingredients
@mysticaltech It is: https://huggingface.co/spaces/Kaikaku/epicure-explorer
@josefchen can it understand that thai fish sauce is equal to parmesan because both are umami? That should be emergent in a recipe model.
@josefchen Is it going to be available at Hugging Face?
