Modular Platform Files

In MAX 26.7, we've added Muse Glimmer and MiMo-V2.6-Flash, plus video input for Qwen3.5 and Qwen3-VL. The new Decisions API returns probabilities for classification, scoring, and yes/no questions without generating text, and works with every text-generation model. Speculative decoding now applies each request's sampling settings to DFlash, DFlash2, and DSpark draft tokens, and MAX can dynamically select the verification width that delivers the best measured decode throughput. We've also made Jenga the default KV cache, added prefix caching for Qwen3.5-family and Inkling models, and enabled tiered KV connectors for recurrent-state models. Plus continued progress on the eager execution API, including your custom ops—eager execution shares a single code path with the graph compiler, so the code you run in eager mode is the same as what you compile and ship. This release also brings substantial performance improvements for models on NVIDIA B200s AMD MI355 GPUs.

With Mojo 1.2, we've added pattern matching for literals, tuples, structs, and more. The same patterns work in match, if, and while statements, and in destructuring assignments. More than 20 std.math functions now share one Mojo implementation across CPU and GPU, bringing functions such as atan(), cbrt(), erfc(), lgamma(), and the Bessel functions to accelerators. Compilation for large packages is much faster, and jemalloc makes large GPU kernel files compile roughly 10 to 13 percent faster on Linux. StringSpan is now immutable, and this release removes MutStringSpan, MutStringSlice, and the mut parameter on StringSpan. And much more.

For all the details, check out the full MAX release notes and Mojo release notes.