Modular 26.3 is out! Here’s what’s new.
Mojo 1.0 is in beta!
Beta 1 is the first step toward finalizing 1.0 later this year. The beta is what we believe is a feature-complete Mojo 1.0 language, with polishing still ahead before final release.
In 26.3, we’re shipping safe closures with new capturing syntax, conditional conformance to traits, and major improvements to variadics. Full details in the Mojo changelog.
Mojo also has a new dedicated home at mojolang.org, consolidating all language docs in one place.
Video generation in MAX
Wan 2.2 support is coming soon for enterprise workloads on Modular Cloud. Text, audio, image, video: one platform, every modality. Request a demo.
Unified, distributed-aware tensor
Writing tensor-parallel code gets a lot easier in 26.3. max.experimental now ships a distributed-aware Tensor type, multi-device compilation, and the collective ops needed for tensor-parallel code. The API maps cleanly to both PyTorch DTensor and JAX NamedSharding patterns. We’ve already built a multi-GPU Gemma3 pipeline on top of it. Docs.
Get started
uv pip install --upgrade modular
MAX changelog | Mojo changelog
Read the full release blog post here.
We’d love to hear what you build with 26.3. Drop your questions, feedback, and projects in the replies.
with hand-written CUDA and Metal GPU kernels, training GPT-2 124M end to end in pure Mojo. On an M4 Max the Metal path runs at 498.92 ms/step in bf16, 1.72x faster than PyTorch MPS (llm.c has no Metal port, so MPS is the baseline there). On a GB10 it reached bf16 parity with llm.c CUDA. Comptime vendor dispatch and MAX’s