Meta's Muse Spark gives Zuckerberg a credible seat at the frontier AI table for the first time

Mark Zuckerberg's bet on a complete rebuild of Meta's AI organization may be paying off, at least in the early going. Wired's analysis of Muse Spark, published the day of its April 8 launch, concluded that the model represents a meaningful upgrade over Llama 4 — the April 2025 release that was widely viewed as disappointing by the AI community — and places Meta's offering credibly among the top-tier models from OpenAI, Anthropic, and Google for the first time.

The model is notably closed source at launch, a departure from Meta's previous open-weights strategy with Llama. Zuckerberg has indicated that open-source versions of future Muse-family models will eventually follow, but the initial closed release reflects a recognition that frontier-level models require significant infrastructure commitments that may be difficult to sustain if weights are freely downloadable immediately.

The move mirrors a broader industry debate about when and how to release powerful models openly.

According to Wired, benchmarks suggest Muse Spark outperforms some recent offerings from OpenAI, Anthropic, Google, and xAI on specific tasks, though comparisons across frontier models are highly dependent on task type. Meta described the model as having been designed with particular depth in health reasoning, coding, and visual STEM — areas where the company sees product traction.

The health angle is a notable strategic choice. Meta says it worked with medical professionals to curate training data that makes Muse Spark's health guidance more accurate and comprehensive than prior models.

Given the potential scale of a Meta health AI deployed across WhatsApp and Facebook — both of which will receive Muse Spark rollouts in coming weeks — this capability will face substantial real-world scrutiny regarding safety and accuracy.

Read the original reporting at Wired.