Meta AI Updates: October 3, 2026

1. Meta Launched Muse Gadgets, Open-Source SDKs for Building Hardware Around Its Muse Agent

Muse Gadgets. Meta Superintelligence Labs released Muse Gadgets, an open-source project that lets developers build their own devices that connect to the Muse agent using off-the-shelf boards such as the ESP32 and Raspberry Pi. The project includes an ESP32 SDK and a Linux SDK under the Apache 2.0 license, published at facebookincubator/muse-gadget-sdk on GitHub, with reference builds for boards including the Waveshare ESP32-S3 AMOLED, M5Stack StickS3, Raspberry Pi 5, and Seeed reTerminal E1002 e-ink display. Meta also built Muse Home Link, a USB-C-powered device that lets Muse control smart speakers, TVs, and other devices on a home network through local HTTP APIs, and is giving 5,000 units to US Muse subscribers for free, one per subscriber, while supplies last. Source

2. Meta Launched Ray-Ban Meta Gen 3 Glasses in India and Said Muse Is Coming to Them

Ray-Ban Meta. Meta launched Ray-Ban Meta (Gen 3) AI glasses in India starting at INR 44,300, with a customizable action button for Meta AI, a 6-mic array that Meta says removes more than 90 percent of background noise, a 12 MP camera with 3K video, and up to 9 hours of battery life. The screenless Ray-Ban Meta Audio model, which weighs 43 grams and runs up to 12 hours plus 48 hours from its case, is coming soon. Meta also said the Muse agent will reach the glasses in the coming months, so it can act hands-free on what the wearer is looking at, such as products on a shelf, fliers, or lists. Source

3. Meta Engineers Open-Sourced a TLX Jagged Flash Attention Kernel That Beats FA4 on Blackwell

TLX. Meta engineers explained how they built Jagged Flash Attention (JFA), the attention kernel inside Meta’s Generative Ads Model (GEM), for NVIDIA B200 GPUs using TLX (Triton Low-level Extensions). JFA runs attention directly on variable-length user sequences packed end to end, which avoids padding that can waste up to 50 percent of compute. In bf16 on the jagged shapes GEM uses, the kernel runs about 13 percent faster than FlashAttention-4 (May 2026 version) on the forward pass and about 50 percent faster on the backward pass, with roughly a third as much code as FA4’s ~10K-line CuteDSL implementation. The code is in the facebookresearch/ads_model_kernel_library repository. Source