Mojo 1.0 compiler and tooling move to Apache 2.0 as Modular expands its AI platform across Trainium, TPUs, and Qualcomm accelerators
- Mojo 1.0 is fully open source under Apache 2.0, including its compiler and tooling; Modular says developers can extend the language and port it to new platforms.
- Modular Cloud is generally available with OpenAI-compatible shared endpoints billed per token and isolated dedicated deployments on Modular or customer compute.
- The Modular Platform now supports AWS Trainium, Google TPUs, Qualcomm Cloud AI 100 Ultra, and Qualcomm Dragonfly alongside CPUs and NVIDIA and AMD GPUs.
- Modular says it is working with the Microsoft Windows team on native Windows support for Mojo, which previously supported Windows only through WSL.
- The MAX license no longer restricts device usage, and Modular plans to make MAX source-available under an open alliance program.
Hacker News opinions
I saw the Phoronix write-up. This is Chris Lattner's project, and his LLVM and Swift background makes it worth watching.
I'd add Swift for TensorFlow to that list too. That effort is relevant context for Mojo's AI angle.
I'm still unsure whether Mojo has meaningful traction or a clearly defined problem. It barely appears in normal programming-language discussion, then resurfaces every year or two.
I think the closed compiler kept people from trying it. Most developers now expect language development to happen in public, and Mojo was an outlier.
I would not judge a language by HN attention. I do not want to trace a GPU runtime error across CUDA, CUTLASS, cuDNN, and PyTorch if Mojo can simplify that stack.
For me, the appeal is a systems language that feels more natural than Rust. GPU programming is useful, but I would also use it for general systems work.
The benchmark numbers have not convinced me. Some GEMM results appear to compute only every eighth value, and the verifier does not validate the whole output, so near-peak TFLOPs can be misleading.
I waited for open source before trying Mojo, and I have started learning it this week. The Python-like syntax, comptime, tile support, GPU work, and puzzles have been enjoyable so far.
The delayed open-source release probably burned much of Mojo's early traction. I still like the idea and will see whether it fits anything I build.
For hobby projects, open source was a requirement for me. Mojo now joins Rust as a language I can seriously consider.
I worry that Qualcomm open-sourcing Mojo after buying Modular means it may extract the cloud business and let the language fade away.
Open sourcing was on Mojo's roadmap before Qualcomm entered the picture. I would not read the acquisition alone as evidence that Mojo is being abandoned.
I was interested when Mojo was presented as a Python superset, but I am more interested now that it has moved away from that. The compiler being open source makes experimenting worthwhile.
I use Rust for work I once wrote in Python, but I do not enjoy Rust for numerical code. Mojo 1.0 still was not ready to port my type-driven DICOM anonymizer, though its numerical side looks promising.
I am interested in the MLIR design: compiler optimizations can live in libraries rather than only in the compiler. Reactant.jl is exploring related ideas for higher-level Julia abstractions.
I am waiting for Phase 2, especially algebraic data types and pattern matching. The current Variant-based tagged unions are verbose and lack exhaustiveness checking and typed error handling.