A new nightly version has been released!
See the quickstart guide for installation instructions: Quickstart | MAX
MAX changelog updates:
- VLM tokenizers can now cache preprocessed media, so an image or video resent
on a later conversation turn skips the resize, rescale and patchify (and for
video, the whole decode) instead of redoing it. Keyed on the same
raw-encoded-bytes digest the vision encoder cache uses, and bounded by host
bytes rather than entry count:--max-vision-preprocess-cache-bytesand
--max-video-preprocess-cache-byteseach default to 10 GiB, their combined
size is capped at a quarter of the memory the process may use (a cgroup grant
where there is one), and0disables either. The budget is a ceiling rather
than a reservation – the cache grows into it and evicts to stay under it –
and on a host with less than 80 GiB the cap scales both down proportionally
rather than overcommitting. Entries unused for
--max-media-preprocess-cache-idle-seconds(default 300,0disables) are
dropped on the next cache lookup or insert, so a burst of distinct media does
not hold host memory for the life of the process. Enabled for Gemma 4 images
and video, Kimi K2.5 images, and Qwen2.5-VL and Qwen3-VL-MoE images.
Mojo changelog updates:
Raw MAX diff: https://github.com/modular/modular/compare/9bbc45645af8504606265bfd6918ee3d3c3cb817...9cabc22e82c42542b712a9758576184e1693f74b)>
Current Mojo changelog: https://github.com/modular/modular/blob/main/mojo/docs/nightly-changelog.md