Meta AI Agent tests thinking mode

Meta is finally adopting the now-standard speed-vs-depth split: instant responses for casual use, slower reasoning when the prompt deserves it. The safest read is that Thinking is a routing layer on Llama 4, but Meta hasn’t confirmed whether there’s a distinct reasoning model behind the switch.

• –Meta’s own launch materials say the app is built with Llama 4, so that remains the only confirmed backend.
• –Reddit speculation points to a reasoning-tuned Llama 4 variant, with some users guessing Meta may be testing its rumored “Avocado” stack.
• –A slow mode makes sense for math, planning, and multilingual edge cases where a few extra seconds can buy noticeably better answers.
• –The bigger story is UX, not branding: Meta is normalizing a speed-vs-depth choice for mainstream users, which makes the assistant feel closer to ChatGPT and Gemini.