Artificial intelligence news.
Straight from the original announcements.

Modular Rebuilds MAX Module for Faster Development and Inference

November 19, 2026By the AI News Bites Editorial TeamSource: Modular
Modular Rebuilds MAX Module for Faster Development and Inference
Image: Modular

Modular has rebuilt the Module API in its MAX AI framework, combining faster eager execution during development with compiled production inference. The redesigned API also adds support for distributed computation and uses a simpler, PyTorch-like syntax.

Key points
  • MAX’s Module API has been rebuilt on a new Eager layer.
  • Modules can run eagerly or compile into optimized graphs.
  • The redesign supports data, pipeline and tensor parallelism.
  • The legacy implementation will be removed in Q1 2027.
  • Modular says 2027 plans include quantization, training and multi-node support.

A new layer over MAX’s graph stack

MAX is designed for production inference, with Modular highlighting its compilation stack as a central advantage. That stack combines cross-architecture support, optimized Mojo kernels, graph compilation and execution that can fuse kernels and plan memory.

The previous Module was built directly on MAX’s graph stack. According to Modular, that graph-forward design delivered strong performance but made the developer experience more difficult. The company has now introduced an Eager layer on top of the graph stack and rebuilt Module on that foundation.

Eager execution starts computing results immediately instead of requiring expensive graph compilation first. Modular says this can speed up unit tests, debugging and development. The company also says that when developers first run Gemma 4 with the new Module, the first token appears much more quickly.

Eager and compiled execution in one API

The redesigned Module remains focused on production inference. Its built-in compile method traces a model and compiles it as a graph, allowing MAX to optimize computation using information about the whole model.

Modular says the graph compiler performs operation fusion to reduce launch overhead, constant folding and other initialization-time lifting to avoid repeating constant computation, and execution and memory planning to reuse buffers rather than allocating new ones. It also supports shape inference and reasoning about dynamic shapes.

Model authors can therefore call a Module eagerly during development or compile it before production use. The new design also follows familiar PyTorch-like patterns, while simplifying them. Developers do not need a super().__init__() call, and can use plain Tensor objects instead of nn.Parameter.

Distributed models and simpler framework APIs

The new Module supports distributed computation across multiple devices. MAX’s lower-level abstractions can express data parallelism, pipeline parallelism and tensor parallelism. A Tensor remains a logical N-dimensional array, but it can now live on more than one device, with operations following defined sharding rules.

Module is also now a Tree, an API inspired by JAX’s Pytrees. Tree traversal helps collect connected Modules and Tensors, while tree mapping can convert Tensors into graph weights. Modular says these utilities also support serialization and can be used to quantize selected parts of a model, such as expert layers in a mixture-of-experts model.

The redesign removes the symbolic Weight class and graph-construction boilerplate. Default dtype and device contexts can prevent developers from passing the same settings into every layer, while each Module now includes a built-in compile method.

The legacy implementation remains in max.nn.legacy and is scheduled for removal in Q1 2027. Modular recommends its Model Bringup Workflow tutorial for migration and says its 2027 plans include quantization, training and multi-node support.

Related news

Modular details its open AI compiler stack at LLVM meeting
For developersNovember 16, 2026

Modular details its open AI compiler stack at LLVM meeting

Modular presented its end-to-end Mojo and MAX software stack at the 2026 LLVM Developers' Meeting, including a full-day workshop and a technical talk on device graphs.…

Together AI says simpler kernels can beat its megakernels
For developersNovember 16, 2026

Together AI says simpler kernels can beat its megakernels

Together AI says it has replaced its difficult-to-deploy AI megakernels with a more adaptable approach built around CUDA Graphs, PDL, Torch Inductor kernels and…

PyTorch Models Can Now Run on GPUs Inside Web Browsers
For developersNovember 10, 2026, 17:45

PyTorch Models Can Now Run on GPUs Inside Web Browsers

PyTorch models can now run on a device’s GPU inside the web browser through the new ExecuTorch WebGPU backend. The approach uses a single ExecuTorch .pte file, avoids…