●Change8

max/v26.4.0

Breaking Changes
📦 mojo-languageView on GitHub →
⚠ 2 breaking✨ 20 features🔧 10 symbols

Summary

The MAX 26.4 release significantly expands hardware support, bringing Apple silicon GPU serving to more models and introducing new quantization and architecture features for models like DeepSeek and Gemma. The Mojo 1.0 beta 2 release focuses on language stability, removing copy requirements for collections and enforcing explicit standard library imports.

⚠️ Breaking Changes

  • The use of `fn` function declarations in Mojo is now a hard compilation error, replacing the previous behavior where `def` and `fn` were partially unified. Users must ensure all function declarations use `def` or update existing `fn` declarations.
  • Implicit imports of the standard library (`std`) are now an error in Mojo. All standard-library dependencies must be explicitly imported.

Migration Steps

  1. In Mojo, ensure all standard-library dependencies are explicitly imported, as implicit `std` imports are now errors.
  2. In Mojo, rename usages of `mojo package` to `mojo precompile`.
  3. In Mojo, update file extensions from `.mojopkg` to `.mojoc`.

✨ New Features

  • Apple silicon GPU serving is now available for almost all compatible MAX models, including M3 and newer devices running Llama and Qwen families.
  • Added support for Tencent Hunyuan Hy3-preview model.
  • Added support for LiquidAI LFM2 model.
  • Added support for GLM-5/5.1 with FP8/NVFP4 weights.
  • DeepSeek V2 and V3 can now run on a single GPU, with V3 supporting block-scaled FP8 quantization.
  • Added long-context sparse MLA support for DeepSeek-V3.2.
  • Gemma 4 gains native FP8 attention with an FP8 KV cache on B200, doubling KV cache capacity while matching bf16 accuracy.
  • Gemma 4 now supports NVFP4 quantization.
  • Gemma 4 now supports structured output with grammar enforcement.
  • Gemma 4 now supports MTP speculative decoding.
  • Gemma 4 31B variant now supports multi-GPU.
  • MiniMax-M2 gains MXFP4 quantization.
  • MiniMax-M2 gains combined tensor-parallel attention and expert-parallel MoE (TP+EP).
  • Enhanced Prometheus metrics.
  • Improved kernel performance.
  • New benchmark datasets added.
  • Numerous Python API upgrades.
  • Mojo 1.0 beta 2 removes the requirement for collection elements to be `Copyable`; `List`, `Deque`, `LinkedList`, `InlineArray`, `Dict`, and `Set` now accept move-only elements.
  • Mojo now supports trailing `where` clauses on struct declarations and `comptime` alias declarations.
  • Mojo `String` and `StringSlice` gain Unicode-aware subscripts.

Affected Symbols

ListDequeLinkedListInlineArrayDictSetStringStringSlicemojo packagemojo precompile