.png)
Modular 26.6 is our first release since open sourcing the Mojo compiler at ModCon. With Mojo 1.1, we’re taking the next step in that journey by opening the compiler to external contributions and moving more of our development into the GitHub repository.
MAX 26.6 expands to support audio generation, new frontier models, and delivers significant performance improvements across NVIDIA and AMD GPUs. We’re also actively working on our ecosystem alliance program and plan to launch it toward the end of 2026 (reach out to alliance@modular.com for more information).
Mojo 1.1: the compiler now accepts contributions
At ModCon in August, we announced the open sourcing of the Mojo compiler under the Apache 2.0 license. Making source code available was the first step and today we’re happy to announce that we’re accepting contributions to the Mojo compiler!
This was one of the top requests we received after open-sourcing the Mojo compiler, and we appreciate your patience while we got the right infrastructure in place. To make it even easier to follow work in progress and collaborate directly with the Modular team, we've migrated our internal issues to public GitHub issues.
Mojo 1.1 builds on the foundation of Mojo 1.0 with a number of compiler, language, and developer experience improvements:
- Compilation time speedup: Compilation of implicit conversions and use of t-string literals has improved.
- LSP improvements: Compiler suggestions for common off-by-one typos.
- Contextual type inference: Allows declarations such as
DType.float32to be shortened to.float32when the type can be inferred. - OSS directory organization: A unified top-level Mojo directory containing the compiler and standard library, along with a new guide to working with Mojo in open source.
See the Mojo 1.1 changelog for the complete list of language and tooling improvements.
MAX 26.6: new modalities and models with faster performance
MAX development continues rapidly and we are committed to making MAX source available. In 26.6, we expanded MAX to support audio output modality, added additional frontier models, and made significant performance improvements across NVIDIA and AMD GPUs.
Audio model generation support: MAX now supports audio generation, expanding beyond text, vision, and image generation. The audio_generation pipeline launches with support for MiniMax-Music3 (MiniMaxAI/MiniMax-Music3), a text-to-music model that generates 44.1 kHz stereo audio from a style prompt and lyrics. Audio generation is available through the /v1/audio/speech and /v1/responses API endpoints, with controls for duration, denoising steps, guidance scale, and output format.
Model architecture support: MAX 26.6 adds support for new frontier model architectures while expanding capabilities across existing model families. This includes support for Thinking Machines’ Inkling (InklingForConditionalGeneration) architecture, along with improvements to the Kimi, Qwen, Gemma, and GLM model series. Speculative decoding techniques (such as DFlash and DSpark) are now supported within MAX, along with improved tool calling and datatype support.
Performance improvements: MAX improves performance across hardware and workloads, delivering broad gains across representative benchmarks. On NVIDIA B200, these optimizations deliver up to 4.8× faster Gemma 4 decode attention, up to 7.9× faster MoE routing, and up to 16× faster low-batch NVFP4 quantization. On AMD MI355, the optimizations deliver up to 6.6× faster performance for decode attention projections, while and the MXFP8 QKV projection is up to 4.8× faster than the previous release.
For the complete list of updates, including additional model support, serving improvements, and speculative decoding controls, see the MAX 26.6 changelog.
Get started with 26.6 and Mojo 1.1
Install or upgrade to get started in minutes:
As announced in 26.5, the modular package is retired in this release. Instead install max["all"] for everything, or max["serve"] if you just need an endpoint, or max["benchmark"] if you just need to benchmark an endpoint (use max-serve or max-benchmark with conda).
The compiler is open to your patches now, and the standard library and compiler issue trackers show what our teams are working on, so the contributing guide and the open issues are the place to start if you want to pick something up. If 26.6 breaks something for you, file it there too. For questions about the release, the Modular forum is the best place to get an answer from the team, and our Discord server is the best place for real-time conversation with other developers building on Mojo and MAX.
