MAX Nightly 26.2.0.dev2026020805 Released

Nightly
1

:astronaut: A new nightly version has been released! :astronaut:

See the quickstart guide for installation instructions: Quickstart | Modular

MAX changelog updates:

  • Enabled overlap scheduling for select model architectures like
    LlamaForCausalLM_Legacy by default. This optimization reduces CPU overhead
    by overlapping python host code with GPU kernel execution. This optimization
    is currently incompatible with some features such as structured outputs or cpu
    models. This feature is very experimental! You can forcibly disable it via
    --no-enable-overlap-scheduler --force.

Mojo changelog updates:

  • [KGEN][oss][max][GC][Kernels] Rename TrivialRegisterType to TrivialRegisterPassable.
  • [KGEN][oss][max][GC][Kernels] Rename RegisterType to RegisterPassable.

Raw MAX diff: https://github.com/modular/modular/compare/a968ccc177c8d1b2caca8b4f4dd62cecb75cd726...140159fe1966f226083dc0640fee70422b561083)>
Current Mojo changelog: https://github.com/modular/modular/blob/main/mojo/docs/changelog.md

1 Like
2

I’ve installed pixi on my Linux machine and I therefore love the intergration of having all Rust ,Mojo and Python in the CLI interface. But my main question is specialized on wether there could a possibility to grant an FFI interface on Pixi to a low level Domain specific language specifically P4- Programmable protocol independent packet processor’s to access Mojo’s AI logic directly for network analysis. Please inform me on the intergration coz I would really love to run Mojo logic directly on the NIC level… I’ll appreciate the answer :grin: