Meta AI Updates: May 13, 2026

1. ExecuTorch on Arm CPUs and NPUs gets a hands-on lab walkthrough

Meta. Meta and Arm published a joint PyTorch blog post by Matt Cossins detailing ExecuTorch deployment paths for Cortex-A, Cortex-M, and Ethos-U targets. The post benchmarks OPT-125M on a Raspberry Pi 5 with the XNNPACK backend using KleidiAI microkernels, and walks through INT8 quantization, TOSA lowering, and Vela compilation for Ethos-U85 NPUs. Companion Jupyter labs ship under arm-education/executorch_on_arm_labs. Source

2. PyTorch core clarifies LibTorch ABI Stable vs TVM-FFI roles

Meta. PyTorch core maintainer Jane Xu published a positioning note explaining how LibTorch ABI Stable and TVM-FFI are complementary rather than competing. TVM-FFI targets a minimal, framework-agnostic stable ABI layer, while LibTorch ABI Stable targets existing PyTorch APIs. Both focus on precompiled kernels and exclude JIT paths like Triton or CuTe DSL, with kernel writers, Rust-binding authors, and compiler backends called out as the primary beneficiaries. Source

3. Threads pilots a Meta AI mention feature similar to Grok

Meta. Meta began beta-testing a Threads integration where users with public accounts can mention @meta.ai in a post or reply to get real-time context, trend explainers, and recommendations as public replies. The rollout starts in Malaysia, Saudi Arabia, Mexico, Argentina, and Singapore, with the assistant matching the original post’s language and supporting per-user muting. Meta told TechCrunch the feature ships with stronger safeguards than comparable competitors. Source

4. PyTorch 2.12 GA targeted for May 13

Meta. The PyTorch release team’s key-dates thread reaffirms PyTorch 2.12 GA on May 13, 2026, with RC1 binaries already on the test channel and M5 content finalized on May 8. The release introduces an experimental CUDA 13.2 build and deprecates CUDA 12.8, with CPU, ROCm, and XPU wheels shipping alongside. Source

5. Meta ships expanded parental supervision tools across its apps

Meta. Meta rolled out supervision features giving parents visibility into their teen’s algorithmic recommendations and account activity across Facebook, Instagram, and Messenger. The launch leans on Meta’s AI-driven age inference and content classifiers to enforce teen account defaults, and is positioned alongside the company’s ongoing under-age enforcement push announced earlier this month. Source