Meta Muse Spark: Multimodal Model, Subagents & Closed Release

Instant answers, or a crew of subagents working the problem

Meta described Muse Spark as a natively multimodal reasoning model with support for tool use, visual chain of thought and multi agent orchestration. Meta AI offers it in two modes: an instant mode for quick answers and a thinking mode, which the technical post calls contemplating mode, that can launch multiple subagents in parallel on one question, for example drafting an itinerary, comparing destinations and finding activities as three separate subtasks running at once. The model can also read images, for instance ranking snacks in a photo by protein content, and Meta paired the release with visual coding, Marketplace shopping integration, live camera perception and voice conversation in the Meta AI app. [1][2]

Meta's own reported evaluations put Muse Spark ahead of rivals on health tasks and behind them on coding and abstract reasoning. On HealthBench Hard, an evaluation built with more than 1,000 physicians, Meta reported 42.8 percent, ahead of Claude Opus 4.6's 14.8 percent and just short of GPT-5.4's 40.1 percent by that read. In contemplating mode, Meta reported 58 percent on Humanity's Last Exam and 38 percent on FrontierScience Research. Independent reporting put its SWE-bench Verified score at 77.4 percent and its ARC-AGI-2 score at 42.5, both trailing the coding and abstract-reasoning scores rival frontier models were posting at the time, and Meta's own post acknowledged gaps in long horizon agentic systems and coding workflows. [2][3]

A ground-up rebuild, credited to the team Meta paid billions to get

Meta Superintelligence Labs said it spent the nine months before the release rebuilding Meta's AI stack from the ground up rather than iterating on the existing one, revising model architecture, optimization and data curation in the pre-training stack. Meta's claimed payoff was efficiency rather than a benchmark record: it said Muse Spark reached comparable capability to Llama 4 Maverick using more than an order of magnitude less compute, and it tied the release to continued investment across research, training and infrastructure, including the Hyperion data center. [2]

The lab doing the rebuilding did not exist a year earlier. Meta had formed Meta Superintelligence Labs after paying $14.3 billion for 49 percent of Scale AI and installing its founder, Alexandr Wang, as chief AI officer to run the new group. Muse Spark was that group's first shipped model, arriving roughly ten months after the deal that created it. [3]

The company that published Llama chose not to publish this one

Muse Spark powers the Meta AI app immediately, with a rollout planned to WhatsApp, Instagram, Facebook, Messenger and Meta's AI glasses, and Meta opened a private API preview to select partners rather than releasing weights. Meta said it intends to open source future versions of the model, framing the closure as a present choice rather than a permanent one. [1][2]

That is a reversal fromLlama 4, released a year earlier with published weights, and from the open-weight strategy Meta had followed since the first Llama model. Coverage of the release described Meta's developer base, built on years of open Llama access, reacting to a closed model from the company that had built that access. [3]

Atlas interpretation: The benchmark gaps in coding and abstract reasoning explain why this release was read as a step rather than a leapfrog. The closure is the more consequential fact: Meta had spent two years building its identity in AI around giving weights away, and Muse Spark is the point where that identity stopped being unconditional. A promise to open source later does not undo shipping the first model closed; it just says the company is treating openness as a decision to make release by release rather than a policy it is bound to. [1][3]