An integrated platform for the “running” stage of AI models — inference. Built around its own programming language, Mojo, and its inference engine, MAX, it covers everything from writing GPU kernels to serving models and delivering them from the cloud as a single continuous stack. Its biggest selling point is that it supports both NVIDIA and AMD GPUs, delivering fast inference without rewriting code for each hardware target. The company was founded in 2022 by Chris Lattner, creator of LLVM and Swift, and Tim Davis, formerly of Google. An agreement to be acquired by Qualcomm was announced in June 2026, and the acquisition closed on July 29, 2026. Mojo, MAX, and Modular Cloud all continue as products and brands.
Key Features
- Mojo — a fast language with Python-like syntax: It reads much like Python, yet can express low-level work such as GPU kernels. You can move back and forth with existing Python assets and replace only the parts that need performance
- MAX — a hardware-agnostic inference engine: The
max servecommand brings a model up as a local endpoint. It supports NVIDIA (B200 / H200 / H100) and AMD (MI355X / MI325X / MI300X), and is designed with CPUs and NPUs in view as well - OpenAI-compatible API: Endpoints are OpenAI-compatible and can be called directly from the OpenAI Python SDK. Beyond text-to-text, image and video inputs are supported, so an existing application can migrate by simply pointing at a different endpoint
- Benchmarking built in:
max benchmarkmeasures performance across different datasets and concurrency levels, letting you confirm the effect of tuning in numbers as you go - Choice of deployment models: Modular’s own cloud, BYOC (Bring Your Own Cloud) inside your VPC or on-premise, or fully self-hosted. This makes adoption practical even when data cannot leave your environment
- Open source and container distribution: The core is published on GitHub and free to use under a Community License. The distribution container is lightweight, under 700MB
Pricing
| Plan | Billing model | Main features |
|---|---|---|
| Community (self-hosted) | Free | Available under an open source license. Community support via Discord and GitHub |
| Modular Cloud | Usage-based (per token for shared endpoints, per GPU hour for dedicated endpoints) | Always-on inference infrastructure, usage metrics, SOC 2 Type 2 certified. Free tier available |
| Bring Your Own Cloud | Per minute of deployed GPU capacity | Runs in your own cloud or on-premise, with data staying inside your VPC. Custom APIs and forward-deployed engineer support included |
| Enterprise / Self-Hosted | Contact for pricing | Runs on AWS, GCP, Azure, or Oracle, with SLA-backed support |
Per-model rates for shared endpoints are published — for example, DeepSeek V4 at $1.74 input / $3.48 output per 1M tokens, and the lighter DeepSeek V4 Flash at $0.14 input / $0.28 output. Volume and committed-use discounts are also available.
Pricing is current as of August 2026. Please check the official site for the latest details.
Pros & Cons
✅ Pros
- Spans NVIDIA and AMD without code changes, reducing dependence on a single GPU vendor
- The OpenAI-compatible API keeps migration costs low for existing applications
- Supports deployment in your own cloud or on-premise, so it works for workloads where data cannot leave your environment
- The core is open source, so you can start with free self-hosting
- Being part of Qualcomm adds engineering resources and hardware-integration backing
⚠️ Cons
- This is inference infrastructure, not a no-code chat service; knowledge of GPUs and containers is assumed
- Mojo is a relatively new language, with fewer libraries and case studies than the Python ecosystem
- Production pricing depends on your configuration, and quotes often require contacting sales
- Support on consumer-grade GPUs comes with limitations
- Coming right after the acquisition, product lines and delivery models may still change
Comparison with Similar Services
| Criteria | Modular (MAX) | NVIDIA TensorRT-LLM | vLLM | Together AI |
|---|---|---|---|---|
| Positioning | Inference platform + own language + cloud | Inference optimization for NVIDIA GPUs | Open source inference server | Managed inference cloud |
| Supported hardware | NVIDIA / AMD and others | NVIDIA only | Mainly NVIDIA (some others) | Managed by the provider |
| Self-hosting | Yes (free license available) | Yes | Yes (fully free) | No (API only) |
| Dedicated language | Mojo | None (CUDA / C++) | None (Python) | None |
| Typical users | Companies and developers running their own inference stack | Engineers optimizing NVIDIA environments | Researchers and individual developers | Developers who prefer not to own infrastructure |
Who Is It For
- Infrastructure and MLOps engineers who want to build and operate their own LLM inference stack
- Organizations struggling with NVIDIA GPU availability or pricing who want AMD and other GPUs as options
- Companies that cannot send data to external APIs and need inference inside a VPC or on-premise
- Developers who want to reach kernel-level optimization but would rather avoid the learning curve of C++ / CUDA
- Teams that want to keep their existing OpenAI-compatible clients while switching only the execution environment to their own
Summary
Modular is a platform that ties language (Mojo), engine (MAX), and cloud into one flow, aiming to decouple AI inference from any single GPU vendor. It is not a chat tool you can casually pick up; it is a tool for the side that operates inference itself. The Qualcomm acquisition has closed, but Mojo, MAX, and Modular Cloud have been explicitly stated to continue, and the commitment to an open, heterogeneous hardware ecosystem is being maintained. A realistic path is to start with self-hosting under the free Community license, then consider Modular Cloud or BYOC once you have a clearer sense of scale.