Meta AI gains task automation with Muse Spark 1.1 amid benchmark-gaming doubts

Muse Spark 1.1 marks Meta's shift from a conversational assistant to an agentic one that acts proactively. The model can make plans, connect to email and calendar applications, build presentations, and follow through on complex requests without continuous re-prompting — capabilities Meta frames as progress toward CEO Mark Zuckerberg's 'personal superintelligence' vision. Example use cases include planning renovations or building training schedules by learning user context over time.

Meta paired the assistant upgrade with a consumer AI product: a Seller App that brings photo-to-listing AI, powered by Llama 4, to Facebook Marketplace, letting users generate listings from a single photo. The broader Meta Superintelligence Labs effort continues expanding with Muse Spark, Muse Image, and a previewed Muse Video capability.

The rollout, however, arrived under a cloud of benchmark skepticism. Critics revisited Meta's April 2025 LMArena scandal, when Llama 4 Maverick fell from 2nd to 32nd after a specially tuned version was submitted for benchmarking — WION explicitly asked whether Meta is 'doing the same with Muse Spark 1.1.' That history makes Meta's capability claims a harder sell than rivals'.

Infrastructure context reinforces Meta's all-in posture: the company plans to deploy 7 gigawatts of compute in 2026, doubling to 14 gigawatts in 2027, with its in-house 'Iris' AI chip entering production in September. Skeptics on Reddit treated the hardware buildout as significant but secondary to model capability. What to watch: whether independent evals validate Muse Spark 1.1's agentic claims, and whether the task-automation features roll out beyond select markets.