**What current Mesa offers (main at c0235eff, Nov 7, 2026):** - Teflon has three back ends: etnaviv (VeriSilicon VIPNano NPUs on Amlogic A311D and NXP i.MX8M Plus), rocket (the Rockchip RK3588 NPU) and ethosu (Arm Ethos-U; its driver is in the tree, but docs/teflon.rst doesn't list it yet). - The build docs say Teflon works only with those drivers. - The models tested are UINT8 image networks: MobileNet V1 and V2, and SSDLite MobileDet.

**Why that doesn't fit the BeagleY-AI (AM67A, also called J722S):** - Its AI accelerators are TI's C7x DSPs with MMA, driven through TI's TIDL. No Mesa driver exists for them, so Teflon has nothing on this board to delegate to. - Its GPU, an Imagination BXS-4-64, is used by Mesa's PowerVR driver for graphics, not by Teflon. - Using Teflon here would mean writing a new Gallium ML driver for C7x/MMA, a large project that depends on how much of the hardware TI documents.

**Whisper:** it's a transformer (attention, matrix multiplication, layer norm, softmax). Teflon has been tested on convolutional image networks only, so even on a supported NPU, Whisper through Teflon is unproven. The practical route on this board is to run whisper.cpp (or similar) on the four Cortex-A53 cores, with the tiny or base model. That needs a C/C++ build in the BQ2 seed, plus audio input. The image has no ALSA userland, and no microphone path has been checked yet.

**Weighed against cleave's plan:** not now. The board's boot, Wi-Fi and HTTPS bring-up come first: the image is still awaiting a boot test, and HTTPS waits on bootstrap-qcow2 ticket d78258f03b. After that, Whisper on the CPU is the doable first step. Teflon would need the C7x driver first.