Pixtral 12B review: vision tests with Open WebUI

Uploaded September 18, 2024, this review tests Mistral AI's Pixtral 12B, with a model label ending in 2409. The speaker describes an Apache 2.0 license, a 12 billion parameter decoder based on Mistral NeMo, and a 128,000-token context window with support for multiple images and varying image dimensions. These details reflect the release discussed in the video.

The demonstration connects a locally hosted Open WebUI frontend to an OpenAI-compatible API on Vultr, the video's sponsor. Inference runs on a cloud NVIDIA L40 with 48 GB of VRAM. This is a self-hosted frontend connected to remote compute, rather than an offline demonstration or a guide to run models locally.

The speaker reports good results for describing a llama, identifying Bill Gates, reading a CAPTCHA, explaining a meme and converting a table screenshot to CSV. The phone-storage test reads several values but misinterprets the cloud download icon. A simple sketch produces working HTML; a more detailed webpage recreation misses images and gets the layout wrong.

Text tests expose limitations: the generated Tetris program raises an attribute error, and the model miscounts the letter R in strawberry. It also fails to extract a URL from a QR code. The Waldo test needs follow-up prompts before the speaker accepts its coordinates. These examples support the speaker's preference for using Pixtral for vision tasks, but do not establish general accuracy.