Pixtral 12B

Fixed-resolution vision encoders force a resolution/aspect-ratio choice at design time and cannot adapt token budget to the input.

Trains a vision encoder FROM SCRATCH that ingests images at their NATURAL resolution and aspect ratio, so the user chooses how many visual tokens an image consumes. Combined with a 128K-token context window it accepts any number of images.

Outperforms Llama-3.2 11B and Qwen-2-VL 7B, and outperforms the much larger Llama-3.2 90B while being 7x smaller. These are quality/size claims; there is no vision-token-cost or latency measurement in the report.

Technical report, released Apache-2.0. THE IMAGE-SPLITTING IS THE POINT FOR THIS REPO: Pixtral's variable-size patchification makes the token count a per-request operator choice rather than a model constant, so serving cost becomes input-dependent - the Pixtral implementation in vLLM-style stacks allocates vision tokens from the image's own dimensions and the scheduler has to plan for variable vision prefill. This is both more faithful and harder to serve than a fixed 336px/576px grid. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.