# Modular Platform

> A unified AI development and deployment platform built on the open-source MAX framework and Mojo programming language, supporting inference across NVIDIA, AMD, and other hardware.

The Modular Platform is an open-source AI development and deployment stack built by Modular, combining the MAX inference framework and the Mojo programming language into a single unified system. The repository, hosted at github.com/modular/modular, has accumulated over 29,000 stars and is actively maintained with regular releases. The platform targets developers and enterprises who need high-performance, hardware-portable AI inference without being locked into a single GPU vendor.

## What It Is

Modular Platform is a full-stack AI infrastructure solution that spans from low-level GPU kernels all the way to cloud-hosted model endpoints. At its core are two open-source components: **MAX** (an open and extensible framework for building and serving AI models) and **Mojo** (a Pythonic systems programming language designed for AI-era performance on CPUs, GPUs, and ASICs). Together, they form what Modular describes as "a unified stack, from the kernels to the cloud," enabling deployment across NVIDIA, AMD, Google TPUs, AWS Trainium, Qualcomm, Intel, ARM, and Apple Silicon hardware.

## Architecture: From Kernels to Cloud

The repository contains several distinct layers that work together:

- **Mojo compiler** (`/KGEN`) and **Mojo standard library** (`/mojo/stdlib`) — the language runtime and toolchain
- **MAX accelerator library** (`/max/kernels`) — GPU kernel implementations written in Mojo
- **MAX inference server** (`/max/python/max/serve`) — an OpenAI-compatible HTTP endpoint
- **MAX model pipelines** (`/max/python/max/pipelines`) — Python-based model graph definitions
- **Code examples** for both MAX and Mojo

The inference server is OpenAI API-compatible, meaning existing applications can swap in Modular endpoints with a single line change. The self-hosted container is under 700MB and can be installed via Docker, pip, uv, pixi, or Conda.

## Deployment Models

Modular offers three distinct deployment paths:

- **Self-Hosted (Community Edition):** Free and open source under the Apache 2.0 license with LLVM Exceptions. Users deploy MAX and Mojo on their own infrastructure across any supported hardware. Community support is provided via Discord and GitHub.
- **Modular Cloud:** Managed shared and dedicated endpoints hosted by Modular, billed per token (shared) or per GPU hour (dedicated). Includes forward-deployed engineers, auto-scaling, scale-to-zero, and observability tooling. SOC 2 Type 2 certified.
- **Bring Your Own Cloud (BYOC):** Modular's control plane and engineering team deploy inside the customer's own VPC on AWS, GCP, Azure, or Oracle. Data never leaves the customer's environment. Billed per minute of reserved GPU capacity.

## Supported Models and Hardware

The platform supports hundreds of open-source models. The hosted API catalog includes models such as DeepSeek V4, Gemma 4, Qwen 3 series, Llama Guard, NVIDIA Nemotron, MiniMax, Kimi, GLM, and image generation models including FLUX.2 variants. Hardware support spans NVIDIA B200, AMD MI355X, Google TPUs, AWS Trainium, Qualcomm AI100/AI200, Intel, ARM, and Apple Silicon — all running on a single unified software stack.

## Update: MAX 26.5 / Mojo 1.0.0

The latest release, tagged `max/v26.5.0` and published on August 11, 2026, marks a significant milestone: **Mojo 1.0.0**. This release signals production readiness for the Mojo language itself, which powers all of Modular's GPU kernels. The repository was last pushed to on August 25, 2026, indicating active ongoing development. The project accepts community contributions to the Mojo standard library, MAX accelerator library, MAX model architectures, and code examples, though the Mojo compiler itself is not yet open for external contributions.

## Community and Ecosystem

Modular maintains an active open-source community through Discord (discord.gg/modular), a dedicated forum at forum.modular.com, a Meetup group, regular community calls, and a YouTube channel. The GitHub repository has over 3,100 forks and more than 1,100 open issues, reflecting broad developer engagement. The project is licensed under Apache License v2.0 with LLVM Exceptions for the repository code, while MAX usage and distribution are governed by the Modular Community License.

## Features
- OpenAI-compatible inference API
- MAX inference framework for building and serving AI models
- Mojo programming language for GPU/CPU kernel development
- Support for NVIDIA, AMD, Google TPU, AWS Trainium, Qualcomm, Intel, ARM, Apple Silicon
- Self-hosted community edition (free, open source)
- Managed cloud endpoints (shared and dedicated)
- Bring Your Own Cloud (BYOC) deployment
- Text, image, and video generation inference
- Function calling and reasoning support
- Auto-scaling and scale-to-zero
- SOC 2 Type 2 certified infrastructure
- Forward-deployed engineers for optimization
- Custom kernel development in Mojo
- OpenAI API drop-in compatibility
- Container under 700MB for self-hosting
- Hundreds of supported open-source models
- Usage metrics and observability
- Role-based access control and user management

## Integrations
Docker, pip, uv, pixi, Conda, NVIDIA CUDA, AMD ROCm, Google TPU, AWS Trainium, Qualcomm AI100/AI200, Hugging Face, OpenAI API (compatible), AWS, GCP, Azure, Oracle Cloud, C++, Python

## Platforms
API, CLI, LINUX, MACOS, WINDOWS, WEB

## Pricing
Open Source, Free tier available

## Version
max/v26.5.0

## Links
- Website: https://docs.modular.com/
- Documentation: https://docs.modular.com/
- Repository: https://github.com/modular/modular
- EveryDev.ai: https://www.everydev.ai/tools/modular-platform