A New Era at Meta
Muse Spark didn't emerge from Meta's existing AI research apparatus. It was built from the ground up over approximately nine months by a new team operating under Chief AI Officer Alexandr Wang — the former CEO of Scale AI, whom Meta brought in through a $14.3 billion deal in 2025.
Wang's arrival at Meta was itself a signal of intent. Scale AI had spent a decade becoming the dominant infrastructure layer for AI training data — its clients included OpenAI, Microsoft, and the United States Department of Defense. Bringing Wang in wasn't a hire; it was an acquisition of a philosophy. Meta described the restructuring that followed as a "ground-up overhaul" of its approach to frontier AI.
The result is Muse Spark: a natively multimodal, reasoning-capable model built for consumer deployment — and the clearest sign yet that Meta intends to compete directly with OpenAI and Google at the frontier, not just in the open-source ecosystem it built its reputation in.
What Muse Spark Can Do
Muse Spark accepts text, image, and voice inputs natively — it is multimodal from the model level, not via a bolt-on pipeline. At launch it ships with two modes: Instant mode for fast, conversational responses, and Thinking mode for extended chain-of-thought reasoning on complex problems. A third mode, Contemplating — for very long, multi-step reasoning tasks — is in development and expected to ship later in 2026.
Health intelligence is a standout capability. Meta worked with more than 1,000 physicians to curate specialized health training data, and the model can generate interactive visual displays — charts, breakdowns, and summaries — tailored to health questions. This shows up directly in benchmark performance: Muse Spark is first in the world on HealthBench Hard.
The model supports a 262K token context window — substantially larger than GPT-5.4's 128K, though smaller than Gemini 3.1 Pro's 1M. At launch, Muse Spark is available in the United States only, with a global rollout planned in the months ahead. Access is free via meta.ai and the Meta AI app, with no subscription required.
The Benchmark Picture
Meta published benchmark results alongside the launch, and they show a model with clear strengths and one significant gap.
On HealthBench Hard — widely regarded as the most rigorous medical reasoning benchmark available — Muse Spark scores 42.8, ahead of GPT-5.4 at 40.1 and well ahead of Gemini 3.1 Pro at 20.6. It is the first model to claim the top spot on this benchmark.
On vision reasoning (MMMU-Pro), Muse Spark scores 80.5%, placing it second globally — behind Gemini 3.1 Pro at 82.4%, but ahead of the rest of the field.
The notable weakness is ARC-AGI-2, the abstract reasoning benchmark designed to resist pattern-matching from training data. Muse Spark scores 42.5 — a full 33 points behind GPT-5.4's 76.1. That gap is significant: it suggests the model may struggle with novel, out-of-distribution reasoning problems compared to OpenAI's flagship.
Token efficiency is a bright spot. Muse Spark achieves an AI Index score of 52 using only 58 million tokens — roughly half the 120 million tokens GPT-5.4 requires for a comparable score. For inference cost, that matters considerably.
The Open-Source Reversal
The most consequential aspect of Muse Spark's launch isn't a benchmark. It's a broken promise — or at least a pivot.
Meta built its AI reputation on open weights. Llama 2 and Llama 3 were landmark releases that gave researchers, startups, and independent developers access to frontier-class models without the gatekeeper dynamic of proprietary APIs. Mark Zuckerberg himself wrote an essay titled "Open Source AI is the Path Forward," arguing that openness was both ethically correct and strategically sound.
Muse Spark has no public weights. There is no download link, no Hugging Face page, no research release. Meta has said it "hopes to open source future versions of Muse Spark" — but has provided no timeline, no commitment, and no specifics.
The Register summarized the reaction succinctly with the headline: "Meta's new model is as open as Zuckerberg's private school." The criticism landed. Meta's open-source positioning had been a genuine differentiator — and a community goodwill asset. That asset is now at least partially spent.
Whether this is a permanent strategic shift or a temporary delay while the model matures remains to be seen. But for now, Muse Spark sits firmly in the proprietary camp alongside GPT-5.4 and Gemini — the very models Meta spent years positioning itself against.
Access and Availability
Muse Spark is free to use at meta.ai — no account required at launch, though Meta account sign-in unlocks additional features. The Meta AI mobile app also supports Muse Spark from day one.
Integration across Meta's product surface is planned: WhatsApp, Instagram, Facebook, Messenger, and Meta's smart glasses will all gain Muse Spark capabilities on a rolling schedule. The scale of that distribution — Meta's apps reach over three billion people — means Muse Spark could become one of the most-used AI models in the world even without leading every benchmark.
A public API is not yet available. Developers can apply for private preview access, but pricing and availability timelines have not been announced. For builders looking to integrate Muse Spark into products, the wait continues.