A new nightly version has been released! ![]()
See the quickstart guide for installation instructions: Quickstart | MAX
MAX changelog updates:
- Fixed a crash serving Gemma 4, Gemma 3 multimodal, Pixtral, and Qwen 3
embedding with a batch larger than one. Each of these architectures accepted
max_batch_sizeand never handed it toPipelineModel, which left the base
at its default of 1 while the scheduler dispatched full batches. The model
input buffers a batch stages into are sized from that value, so the first
multi-context step overran them and the model worker crashed. The
architectures pass it up now.
Mojo changelog updates:
Raw MAX diff: https://github.com/modular/modular/compare/579c250800dfde52a4d247fafc53bc048c771cff...26cfe94f400b71afd2cccf439fde95488910c48d)>
Current Mojo changelog: https://github.com/modular/modular/blob/main/Mojo/docs/site/nightly-changelog.md