MAX
ModularUnified inference engine built on the Mojo programming language that targets multiple hardware backends (NVIDIA, AMD, Intel, Apple) from a single codebase. Includes a graph compiler that automatically optimizes model execution for the target hardware. Supports PyTorch and Hugging Face models with claimed performance matching or exceeding TensorRT-LLM on NVIDIA GPUs while also running on non-NVIDIA hardware. Founded by former LLVM/Swift creator Chris Lattner. 26K+ stars on the combined Modular repo.
MAX 26.3 (stable, released May 7, 2026; Mojo 1.0.0b1). High-performance inference-serving framework (500+ open models, OpenAI-compatible endpoints) on NVIDIA/AMD/Apple GPUs + CPU. Source largely public (github.com/modular/modular) but under the non-OSI Modular Community License.
Openness
2 high confidence- license
- Modular-Community-License(non-OSI, per-device licensing and a limited redistribution grant, governs MAX/Mojo usage+distribution)
- repo-license
- Apache-2.0-WITH-LLVM-exception(the source in modular/modular)
- source
- public(~450k LOC incl MAX kernels + inference server + Mojo stdlib)
- commercial-tier
- Modular enterprise/managed deployment
Large public codebase, and the repo's own LICENSE is now Apache-2.0 with LLVM Exceptions. The README is explicit that "Modular, MAX and Mojo usage and distribution are licensed under the Modular Community License", which is not OSI approved: read 2026-07-30, it licenses per discrete physical device, grants only a limited right to redistribute certain components, and provides the community variant free of charge subject to withdrawal at any time. So the source is readable and not freely runnable at scale. Marketed as "open"; flag open_washed. Score corrected from 3 to 2 on 2026-07-30. The software ladder has rungs at 1, 2, 4 and 5 and none at 3, so 3/source_available was not a pair any rule could produce; the ladder scores "you can read it and not run it freely" at 2.
- https://github.com/modular/modular recorded 2026-06-04
Modular Community License (non-OSI) governs MAX/Mojo usage+distribution; ~450k LOC public; MAX 26.3 / Mojo 1.0.0b1 (May 7 2026)
- https://docs.modular.com/max/ recorded 2026-06-04
MAX serves 500+ open models with OpenAI-compatible endpoints across NVIDIA/AMD/Apple GPUs
Adoption
2 low confidenceNo published download/user/tokens figure for MAX serving specifically. 26.3k GitHub stars on the modular/modular monorepo (covers MAX+Mojo) is the only signal; capped per stars_fallback and placed at 2 given no verified production-deployment count and a still-maturing serving ecosystem vs the established OSS engines.
- https://github.com/modular/modular recorded 2026-06-04
26.3k stars on modular/modular (MAX + Mojo)
Capability
4 medium confidenceBroad multi-hardware coverage and a high-performance serving stack with custom-kernel programmability; no standardized MLPerf/AA inference-engine submission verified independently, so feature_matrix. Strong (C4) but not a demonstrated throughput-frontier-definer like the C5 anchors.
- https://docs.modular.com/max/ recorded 2026-06-04
500+ models, OpenAI-compatible endpoints, NVIDIA/AMD/Apple GPU + CPU, custom GPU kernels
Unchanged since 2026-07-30 (last edited, not re-checked)