Modular MAX

Modular Platform releases 25.3: Open-source MAX framework with Google Colab support

Modular Platform's 25.3 release introduces a unified pip install modular package, granting access to Mojo and MAX. This release open-sources MAX Kernels and Serving APIs, totaling over 500,000 lines of code. Google Colab support is now available, enabling execution of MAX models, and a simplified community license for MAX and Mojo, aiming to lower entry barriers. This update reflects a commitment to building in the open and putting the community first. For additional details, checkout the changelog .

Date not specified

Modular Platform 25.4: AMD GPU Support Released

We're excited to announce Modular Platform 25.4, a major release that brings the full power of AMD GPUs to our entire platform. This release marks a major leap toward democratizing access to high-performance AI by enabling seamless portability to AMD GPUs. Developers can now build and deploy models optimized for peak performance, with zero reliance on any single hardware vendor—unlocking greater flexibility, lower costs, and broader access to compute. For more details, see the 25.4 changelog and the release blog post .

Date not specified

Modular Platform 25.5 Released — Large Scale Batch Inference & MAX Graph API

Modular Platform 25.5 is here , and introduces Large Scale Batch Inference: a highly asynchronous, at-scale batch API built on open standards and powered by Mammoth . This release also features the open source launch of the MAX Graph API and expanded support for writing custom PyTorch operators directly in MAX. In addition, we’ve made Modular Platform development and deployment easier with optimized Docker containers and new standalone Mojo Conda packages. Check out all of 25.5’s updates with the full MAX and Mojo changelogs .

Date not specified

Modular 25.6 Release: Expanded GPU Support & New Benchmarking Tool

The 25.6 release is a major milestone in our mission to build a unified compute layer for AI. MAX offers a model serving framework that now spans from consumer CPUs and GPUs to the world's most powerful datacenter GPUs, now including industry leading throughput on both NVIDIA Blackwell (B200) and AMD MI355X. We've also added a max benchmark command to our CLI tool that makes it easy to verify all our performance benefits for yourself. Plus, support for Apple Silicon GPUs is well on its way. For the first time, Mojo developers can directly tap into Mac GPUs—allowing you to write GPU algorithms that run unmodified across Apple Silicon GPUs, NVIDIA Blackwell, AMD MI325X, AMD MI355X, and more. On top of that, you can also now pip install mojo as a standalone package for enhanced Python-to-Mojo interoperability. Check out all of 25.6's updates with the full MAX changelog and Mojo changelog .

Date not specified

modular SDK v25.6.1 released

[Release] Update lockfiles to point to latest stable version: 25.6.1

Date not specified

Modular 25.7: Open-Source MAX Python API and Mojo Updates

The 25.7 release delivers a fully open-source MAX Python API, an experimental PyTorch-like Model API for faster, easier development and compilation, and major performance and portability gains including bfloat16 on Grace Hopper and Grace Blackwell. Mojo also sees significant upgrades with expanded Apple silicon GPU coverage and safer GPU programming features like stricter type checking, improved pointers, clearer error messages, and better Address Sanitizer support. Check out all of 25.7's updates with the full MAX changelog and Mojo changelog .

Date not specified

SAFE-DOCS: New indexing tutorial released

BEGIN_PUBLIC [docs] DOCS-65: Adds indexing tutorial END_PUBLIC DOCS_SYNC_SKIP - Adds new code

Date not specified

Modular 26.1 Release: PyTorch-like APIs and Apple Silicon GPU Support

This release graduates the eager-style Tensor and Module APIs out of experimental, offering a PyTorch-like modeling interface for building high-performance GenAI models—simplifying the process to deploy a model in production with MAX. This release also expands Apple silicon GPU support so basic MAX graphs now compile and run on Apple GPUs (LLM inference support is coming in future updates). 26.1 also includes several new Mojo language features, such as compile-time reflection for automatic trait conformance (enabling things like JSON serialization and CLI argument parsing), explicitly destroyed types ("linear types") for compile-time guarantees that values cannot be leaked, and typed errors for better error-handling on GPUs without overhead. Check out all of 26.1's updates with the full MAX changelog and Mojo changelog .

Date not specified

Modular 26.2 Release: Image Generation, DeepSeek Updates, and Language Changes

The 26.2 release brings image generation to MAX with FLUX diffusion models served through a new /v1/responses endpoint implementing the OpenResponses API. This release also delivers major DeepSeek improvements—DeepSeekV3.2 with multi-latent attention, NVFP4 quantization for DeepSeek-R1, and simplified single-node expert parallelism—alongside new model support for Kimi vision-language models and Qwen3-MoE. This release deprecates Mojo's fn keyword, unifying all function declarations into the def keyword, and merges __moveinit__ / __copyinit__ into a unified __init__() with take / copy arguments. New language features include conditional trait conformances via where clauses, t-strings for structured template processing, comptime if / comptime for syntax, a Python-style assert statement, and much more. Check out all of 26.2's updates with the full MAX changelog and Mojo changelog .

Date not specified

stable model updated to 26.2.0

[Release] Update lockfiles to point to latest stable version: 26.2.0

Date not specified