A new nightly version has been released! ![]()
See the quickstart guide for installation instructions: Quickstart | Modular
MAX changelog updates:
-
API server now returns chunked tokens from the model worker, reducing overhead
and significantly improving throughput for small models and decode-heavy
workloads. -
Renamed
max.driver.Tensortomax.driver.Bufferto clarify that it
represents a low-level memory buffer, not a tensor. The
max.experimental.tensor.Tensorclass remains the primary tensor type. -
Fixed max.driver.Buffer.to(stream) to not copy (i.e. return reference to
Mojo changelog updates:
- [External] [stdlib] Rename
char_lengthtocount_codepointsfor consistency (#75473) - [mojo] Add
@align(N)decorator for struct alignment - [stdlib] Reorganize reflection into submodules
- [stdlib] Move
tcgen05module togpu.compute.arch - [stdlib] Remove deprecated GPU compatibility modules
- [stdlib] Add
write_repr_toandWritableto:(Legacy)UnsafePointer,Pointer,ArcPointer,OwnedPointerandAddressSpace - [stdlib] Add
is_struct_typeto reflection module - [stdlib] Remove
BoolableandDefaultableforError. - [stdlib] Move source location APIs to reflection module
- [External] [stdlib] Make LinkedList conform to Writable, Stringable and Representable (#74799)
- [stdlib] Create default implementation for
Hashable - [stdlib] Add default reflection-based
__eq__forEquatable - [stdlib] Remove implicit copyability from InlineArray (#75295)
- [stdlib] Add tests for
struct_field_type_by_nameas type annotation
Raw MAX diff: https://github.com/modular/modular/compare/ef0181e1e96a9bd9162dc7ac93149d6926a182e5...bcd8f0bffebfc6b6369b3613817c83fcec4e853f)>
Current Mojo changelog: https://github.com/modular/modular/blob/main/mojo/docs/changelog.md