Modular

Modular builds MAX, an inference serving framework, and Mojo, a Python-like systems language for GPU kernels, plus a managed Modular Cloud. It is for platform and ML infrastructure teams that want to serve open models on several hardware vendors from one stack instead of being tied to CUDA.

Modular was founded in 2022 by Chris Lattner, the creator of LLVM and Swift, and Tim Davis, and is based in Palo Alto. Its stack has three parts. MAX is a model serving and modeling framework with OpenAI-compatible endpoints that ships as a container under 700 MB and runs on NVIDIA and AMD GPUs and x86 and ARM CPUs; the company also lists TPU, Trainium, Intel, Qualcomm and Apple silicon support. Mojo is a Python-styled systems language for writing portable GPU kernels; Mojo 1.0 shipped on August 11, 2026 in the Modular 26.5 release, and the compiler was open-sourced under Apache 2.0 with LLVM exceptions on August 18. Modular Cloud runs the same stack as shared per-token endpoints, dedicated per-GPU-minute endpoints, or inside a customer's own VPC. Modular raised a $250 million Series C at a $1.6 billion valuation in September 2025, led by Thomas Tull's USIT fund, bringing total funding to $380 million. Qualcomm completed its acquisition on July 29, 2026 in an all-stock deal reported at $3.9 billion, and says Mojo, MAX and Modular Cloud continue as products and that MAX will keep supporting third-party hardware. Named users include Inworld, MiniMax, Hippocratic AI, TensorWave and AWS. The vendor claims up to 2x the throughput of vLLM on the same hardware; those figures have not been independently verified. The main GitHub repository has about 29,900 stars and 3,200 forks.

Category

Infrastructure & Cloud

Pricing

Free, Usage-based, Contact for pricing

Target Market

CTOs, ML Platform Teams, AI Infrastructure Engineers, Enterprise Developers

Deployment

Open-source, Self-hosted, Multi-cloud, Hybrid

Founded

2022

Headquarters

Palo Alto, United States

Team Size

51-200

• ✓
  MAX inference server

  Serves 1,000+ pre-configured models behind OpenAI-compatible REST endpoints from a container under 700 MB.

• ✓
  Mojo 1.0 language

  A Python-styled systems language for writing GPU kernels once and running them across chip vendors, now stable at 1.0.

• ✓
  Open-source compiler

  The Mojo compiler, standard library and MAX kernels are on GitHub under Apache 2.0 with LLVM exceptions, so teams can audit and fork them.

• ✓
  Multi-vendor hardware support

  Runs on NVIDIA and AMD GPUs and x86 and ARM CPUs, with TPU, Trainium, Intel, Qualcomm and Apple silicon also listed, which reduces lock-in to one accelerator.

• ✓
  Modular Cloud endpoints

  Shared per-token endpoints for testing and dedicated per-GPU-minute endpoints for production, with SOC 2 Type 2 certification.

• ✓
  Bring Your Own Cloud

  Deploys the stack in the customer's VPC on their cloud credits, so data never leaves their environment.

• ✓
  Forward-deployed engineers

  Cloud and BYOC plans include Modular engineers who tune throughput and latency for the customer's specific pipelines.

• •
  Cutting inference cost on AMD

  Move open-model serving onto AMD GPUs without rewriting kernels; TensorWave reports 70% total cost savings on AMD with MAX.

• •
  Low-latency voice AI

  Serve speech and LLM models with sub-second time to first token, as Inworld and Hippocratic AI do for real-time conversations.

• •
  Writing portable custom kernels

  Write performance-critical GPU kernels in Mojo once and run them on several vendors instead of maintaining CUDA and ROCm versions.

• •
  Private inference in your VPC

  Run the managed stack inside your own AWS, GCP, Azure or Oracle account to satisfy data residency and compliance policies.

Self-hosting MAX and Mojo is free. Modular Cloud publishes per-token rates for shared endpoints (DeepSeek V4 at $1.74 input / $3.48 output per 1M tokens, Qwen 3.5 9B at $0.17 / $0.25) and bills dedicated endpoints per GPU minute. BYOC and Enterprise are quote-only, and one analysis warns enterprise pricing could change after the Qualcomm integration. Prices checked on modular.com/pricing on October 2, 2026.

This page was written from 8 sources, 5 on domains other than modular.com.

1. 1.modular.com — modular.comvendor
2. 2.modular.com — pricingvendor
3. 3.modular.com — modular 26 5 mojo 1 0 is herevendor
4. 4.github.com — modular
5. 5.sdxcentral.com — modular raises 250m for ais unified compute layer at 16b val
6. 6.runtimewire.com — chris lattner open sources mojo qualcomm modular
7. 7.developersdigest.tech — qualcomm modular acquisition developer guide 2026
8. 8.hn.algolia.com — search