Summary
Fahd Mirza’s weekly AI recap for May 2026 covers the most consequential model releases, infrastructure updates, and industry deals of a particularly busy week. The headline story is Alibaba’s launch of Qwen 3.7 Max and Qwen 3.7 Plus at the Alibaba Cloud Summit — cloud-only, thinking mode only, with open weights expected but not yet released. The APEX competition-math benchmark lead is real, though Mirza notes that a portion of the accuracy gain comes from the model abstaining more on uncertain questions rather than answering incorrectly — a meaningful trade-off between hallucination rate and answer coverage.
Other model releases include ByteDance’s Lens, a 3B parameter unified multimodal model covering image generation, image editing, video generation, video editing, and video understanding in a single checkpoint; Cohere’s Command A Plus (218B MoE, 25B active, Apache 2.0, 128k context); and Nvidia’s Sana WM, a 2.6B open-source world model that generates 60-second 720p video from a single image and camera trajectory on one GPU.
The infrastructure story of the week is multi-token prediction (MTP) landing in llama.cpp mainline — not a fork or experimental branch — with LM Studio 0.14 already shipping an experimental MTP toggle. Because tools like LM Studio and oobabooga build on llama.cpp, this change cascades broadly across the local AI ecosystem. Industry headlines include Anthropic’s deal with SpaceX for 220,000 Nvidia GPUs (immediately doubling token limits on paid plans), Cursor’s separate SpaceX training deal targeting million-token context, and Google I/O’s introduction of parallel sub-agent support.
📺 Source: Fahd Mirza · Published May 23, 2026
🏷️ Format: Roundup
