The MAX 26.4 release brings Apple silicon GPU serving to almost all compatible MAX models, with M3 and newer devices now running Llama and Qwen model families (and others that fit in system memory). We've also added support for new models, including Tencent Hunyuan Hy3-preview, LiquidAI LFM2, and GLM-5/5.1 with FP8/NVFP4 weights. DeepSeek V2 and V3 can now run on a single GPU (with block-scaled FP8 quantization for V3) and long-context sparse MLA for DeepSeek-V3.2. Gemma 4 gains native FP8 attention with an FP8 KV cache on B200 (roughly matching bf16 accuracy while doubling KV cache capacity) and NVFP4 quantization, structured output with grammar enforcement, MTP speculative decoding, and multi-GPU support for the 31B variant. MiniMax-M2 picks up MXFP4 quantization and tensor-parallel attention combined with expert-parallel MoE (TP+EP). Plus much, much more, including enhanced Prometheus metrics, improved kernel performance, new benchmark datasets, and lots of Python API upgrades.
The Mojo 1.0 beta 2 (v1.0.0b2) release makes steady progress towards a stable 1.0 release later this year. It removes the requirement that collection elements be Copyable: List, Deque, LinkedList, InlineArray, Dict, and Set now accept move-only elements. Trailing where clauses are now supported on struct declarations and comptime alias declarations. String and StringSlice gain Unicode-aware subscripts. The fn function declarations are now a hard compilation error (completing the def/fn unification begun in beta 1). Implicit std imports are also an error, making all standard-library dependencies explicit and freeing common names like algorithm and memory for user modules. On the CLI side, mojo package renamed to mojo precompile and the .mojopkg extension renamed to .mojoc (which is also a smaller, faster format). And that's just the start of it.
For much more detail about the MAX and Mojo changes, see the MAX changelog and Mojo changelog.