The 26.2 release brings image generation to MAX with FLUX diffusion models served through a new /v1/responses endpoint implementing the OpenResponses API. This release also delivers major DeepSeek improvements—DeepSeekV3.2 with multi-latent attention, NVFP4 quantization for DeepSeek-R1, and simplified single-node expert parallelism—alongside new model support for Kimi vision-language models and Qwen3-MoE.
This release deprecates Mojo's fn keyword, unifying all function declarations into the def keyword, and merges __moveinit__/__copyinit__ into a unified __init__() with take/copy arguments. New language features include conditional trait conformances via where clauses, t-strings for structured template processing, comptime if/comptime for syntax, a Python-style assert statement, and much more.
Check out all of 26.2's updates with the full MAX changelog and Mojo changelog.