Qualcomm open-sources Mojo 1.0 and MAX to take on CUDA
Weeks after its $3.9B buy, Qualcomm frees the whole stack under Apache 2.0 — Mojo now targets NVIDIA, AMD, Apple and TPUs. Plus: Claude Code's Concise mode.
Copy markdown
The compiler is finally open
Mojo 1.0's compiler, tooling and standard library are now on GitHub (github.com/modular/modular) under Apache 2.0 with LLVM exceptions. You can build it locally with Bazel, patch the stdlib and kernels, and run the full test suite — though compiler PRs aren't accepted until Modular opens them by end of 2026.
One kernel, every accelerator
MAX and Mojo now target NVIDIA, AMD, Apple, Qualcomm's Cloud AI 100 Ultra and Dragonfly, AWS Trainium and Google TPUs, with a claimed 10x cut in the effort to bring up new silicon. The pitch: write GPU code once instead of hand-porting to CUDA for each chip.
Why Qualcomm gave it away
Qualcomm bought Modular — Chris Lattner's company behind Mojo and MAX — for $3.9B this summer and open-sourced the compiler weeks later, a direct swing at NVIDIA's CUDA lock-in. AMD sharing the ModCon 2026 stage signals this is real multi-vendor buy-in, not a one-chip play.
Modular Cloud goes GA
Shared, OpenAI-compatible endpoints with pay-per-token pricing are now live at console.modular.com, alongside reserved dedicated instances. You can hit MAX-served models today without owning any of the hardware.
Elsewhere: Claude Code adds a Concise mode
Claude Code 2.1.237 ships a built-in "Concise" output style that leads with results and drops the preamble, and fixes prompt caching for sessions running behind an LLM gateway or custom base URL.