Overview
The Modular Platform is a unified platform for AI development and deployment whose open-source components live in the modular/modular repository. It has two halves. MAX is the serving side: an inference server that exposes an OpenAI-compatible endpoint, a library of accelerator kernels, and model pipelines expressed as Python-based graphs. Mojo is the language side: a compiler and standard library aimed at writing high-performance kernels and systems code without dropping to C++.
Both halves are developed in the open, and the repository is explicit about what that means for licensing: the repository and its contributions are under the Apache License v2.0 with LLVM Exceptions, while MAX usage and distribution are governed separately by the Modular Community License. Contributions are accepted to the Mojo standard library, the MAX accelerator library, MAX model architectures, examples and docs; the Mojo compiler itself is not yet open to contributions.
In practice you pick an entry point rather than adopting the whole platform. Teams serving models start from the MAX quickstart and run a model behind the OpenAI-compatible endpoint; teams writing kernels start from the Mojo quickstart and the standard library. The repository lays these out directly — the Mojo compiler and stdlib under /Mojo, the accelerator library under /max/kernels, the inference server under /max/python/max/serve, and the model pipelines under /max/python/max/pipelines — with runnable examples on both sides.
What it does
- MAX inference server with an OpenAI-compatible endpoint, so existing clients can serve against it
- MAX accelerator kernel library and Python-based model pipeline graphs, both in the open repository
- The Mojo compiler and standard library for writing high-performance kernels and systems code
- Apache-2.0-with-LLVM-Exceptions licensing on the repository, with MAX usage and distribution under the separate Modular Community License
- Open contribution paths for the Mojo standard library, MAX kernels, MAX model architectures, examples and docs
- Runnable code examples for both MAX and Mojo, plus developer docs for working inside the codebase
Getting started
The two halves have separate quickstarts — start from whichever job you have. Both are documented on Modular's own docs sites.
Serve a model with MAX
The MAX quickstart guide walks through installing the framework and serving a model behind the OpenAI-compatible endpoint.
https://max.modular.com/get-startedStart with the Mojo language
The Mojo quickstart covers installing the toolchain and writing your first program.
https://mojolang.org/docs/manual/quickstart/Read the source you care about
The repository is laid out by component, so you can go straight to the part you intend to use or extend.
/Mojo # Mojo compiler
/Mojo/stdlib # Mojo standard library
/max/kernels # MAX accelerator library
/max/python/max/serve # MAX inference server (OpenAI-compatible)
/max/python/max/pipelines # MAX model pipelines (Python graphs)Check the licence before you distribute
The repository is Apache-2.0 with LLVM Exceptions, but MAX usage and distribution are covered by the Modular Community License — read it if you plan to ship MAX as part of a product.
https://www.modular.com/legal/communityCommands and code are distilled from the project's own documentation — always check the official repo for the latest.
When to use it
- Serve open models behind an OpenAI-compatible endpoint on your own accelerators using the MAX framework
- Write GPU kernels in Mojo instead of C++/CUDA while keeping a Python-facing model pipeline
- Study or extend a production inference stack whose kernels, pipelines and serving layer are all readable in one repository
- Contribute to the Mojo standard library or add a model architecture to MAX
How Modular Platform compares
Modular Platform alongside other open-source serving & deployment tools AI/TLDR tracks, ranked by GitHub stars.
| Tool | Stars | What it does |
|---|---|---|
| Transformers | ★ 165k | Hugging Face Transformers is a Python framework that defines and runs state-of-the-art pretrained models for text, vision, audio, and multimodal tasks, for both inference and training. |
| vLLM | ★ 91.4k | A high-throughput LLM serving engine that uses PagedAttention and continuous batching to serve many requests at once. |
| SGLang | ★ 35.8k | A serving framework for LLMs and multimodal models that boosts throughput by reusing shared prompt prefixes across requests. |
| Modular Platform | ★ 29.7k | The MAX inference framework and the Mojo language, developed in one open repository |
| TensorRT-LLM | ★ 14.6k | NVIDIA's library that compiles LLMs into optimized engines for the fastest inference on its data-center GPUs. |
| OpenLLM | ★ 12.5k | A tool to run any open-source LLM as an OpenAI-compatible API endpoint locally or in the cloud. |
| LMCache | ★ 11.7k | A KV-cache layer that stores and shares cached attention state across engines and requests to cut repeated computation. |
| NVIDIA Triton Inference Server | ★ 11k | A multi-framework model server that runs TensorRT, PyTorch, ONNX, and other models with dynamic batching and concurrent execution. |