Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities.
> #1 on the @ArtificialAnlys Decision Index (75)
> Read a value from the input (text, number or date)
> Find every passage of a kind in a text (spans)
> Locate things in a picture (box for one, boxes for every instance)
> Estimate a number (number with a range)
> Pick every option that applies (choice)
1/ 🧵
https://t.co/I90nRpjq9g - new-ish custom MoE for my fellow MoE lovers
Qwen4_exp architecture, original base is 3.6-35B
Logit distillation from 3.8-27B (xHigh)
MoE with 3.3B params active, 27B total.
2B n-gram table
Pruned experts per layer to get from 35B -> 27B
give it a shot
Claude built a language model that fits on a floppy disk 💀
367K parameters
~435 KB total
Runs on an Intel Pentium
Fable 5.2 wrote the transformer
Opus 5.5 generated ~3,500 training facts
Something like this. I don’t give a shit that it’s AI. These are the types of movies we’ve been missing for 20+ years, in favor of a million remakes and comic book crap
The models then guesses the image in a limited time, if it has some key words correct in the time, you get a point. If it couldn’t guess your image, the model gets a point.
One HTML file. Soon on GitHub
Currently I build a game you can play with your local vision models.
The user needs to draw a image in a empty canvas about a specific topic. The model now gets the canvas every few seconds and has to guess your image.