Hacker News Reader

Downloaded params.json - GeLU & 2D RoPE are used for the vision adapter. The vocab size also got larger - 131072 in size.

Also Mistral's latest tokenizer PR shows 3 extra new tokens (the image, the start & end).

The torrent is 24GB, and I guess the implementation will be up in HF in a few days!

Exciting times!

At 24gb, I wonder how small the quantized models will go.

The name ist pixtral_12b_240910, size is 24gb. My best guess would be it's a multi modal LLM.

It is vision only (not multimodal)

From the 1st link:

Pixtral is out! Mistral common has image support! You can now pass images and URLs alongside text into the user message.

I mean, it supports text+images, not documents or audio

>text+images

So, it's multimodal then.

will this be supported in ollama?

Here is the feature request to support Pixtral on llama.cpp: https://github.com/ggerganov/llama.cpp/issues/9440 - once it's implemented in the upstream, ollama will be able to pick it up.

New Mistral AI Weights