September 2026 Summaries

Modular’s 26.6 release introduces Mojo 1.1 and MAX 26.6, marking further progress toward open development of its AI software stack. Mojo’s Apache 2.0-licensed compiler now accepts external contributions, while internal issue tracking has moved to public GitHub; the release also improves compilation speed, language-server suggestions, contextual type inference, and open-source project organization. MAX 26.6 adds audio generation through a pipeline initially supporting the MiniMax-Music3 text-to-music model, accessible through audio and responses API endpoints with configurable output controls. It also expands support for architectures and model families including Thinking Machines’ Inkling, Kimi, Qwen, Gemma, and GLM, while adding speculative decoding methods and improving tool calling and datatype support. Performance optimizations target NVIDIA and AMD GPUs, with reported gains in decode attention, MoE routing, quantization, and projection workloads. The retired modular package is replaced by MAX installation options tailored to full, serving, or benchmarking use cases, and Modular plans to launch an ecosystem alliance program near the end of 2026.