The model is natively multimodal: it accepts voice, text, and image inputs, though it produces text-only output. It powers Meta AI across the Meta ecosystem, the standalone Meta AI app, Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses, rolling out over the next several weeks.

Three interaction modes ship at launch: Instant (fast, direct answers), Thinking (pauses to reason before responding), and Contemplating (orchestrates multiple parallel reasoning agents for the hardest tasks). Contemplating mode is genuinely novel, no other major model ships this as a native feature today.

Benchmark Comparison: Muse Spark vs GPT-5.4 vs Claude Sonnet 4.6 vs Gemini 3.1 Pro

The numbers tell a clear story. Muse Spark is not the best frontier model overall, it scores 52 on the Intelligence Index versus 57 for both GPT-5.4 and Gemini 3.1 Pro. But it leads on health AI by a 2x+ margin, matches Gemini on token efficiency, and is completely free. For developers and users in health, medical, and wellness use cases, Muse Spark is the current state-of-the-art.

Where Muse Spark Wins

1. Health and Medical AI, By a Wide Margin

42.8 on HealthBench Hard is a landmark score. GPT-5.4 hits 40.1. Gemini 3.1 Pro gets 20.6. Claude Sonnet 4.6 scores 14.8. Muse Spark is not slightly better, it is categorically better for health-adjacent tasks. If you are building a medical AI product, Muse Spark should be your first API call once public access opens.

2. Token Efficiency

Muse Spark completed the full Intelligence Index evaluation using just 58 million output tokens, matching Gemini 3.1 Pro and dramatically below Claude Sonnet 4.6 (157M) and GPT-5.4 (120M). Lower token usage means lower API costs when paid access arrives. For high-volume applications, this matters enormously.

3. Contemplating Mode

Multi-agent parallel reasoning is new. In Humanity's Last Exam (HLE) scores, Contemplating mode beats both GPT-5.4 and Gemini. Meta is betting that orchestrating multiple reasoning agents is the right architecture for the hardest problems, and initial results suggest they are correct.

4. Free Access

Free is not a benchmark. But for the millions of developers who cannot afford GPT-5.4 Pro at $200/month, free frontier-adjacent performance is a real advantage. Muse Spark is the most accessible model near the frontier today.

Where Muse Spark Loses

1. Abstract Reasoning: A 34-Point Gap

ARC-AGI-2 score of 42.5 versus GPT-5.4 at 76.1 and Gemini at 76.5. Abstract reasoning is the core test of general intelligence, the ability to extrapolate patterns never seen before. Muse Spark lags significantly. This is not a minor shortcoming; it is a fundamental capability difference.

2. Coding: 16 Points Behind GPT-5.4

Terminal-Bench 2.0 score of 59.0 versus GPT-5.4 at 75.1 and Gemini at 68.5. For software development and code generation, Muse Spark is not the right tool. GPT-5.4 and specialized coding models remain the professional choice.

3. Agentic Tasks: 228-Point ELO Gap vs GPT-5.4

GDPval-AA puts Muse Spark at 1,444 ELO, versus GPT-5.4 at 1,672 and Claude Sonnet 4.6 at 1,607. For autonomous AI agents doing real-world tasks, Muse Spark is a tier behind. This matters for AI agent applications, which are the dominant product direction across the industry in 2026.

4. No Public API Yet

A private API preview is available to select enterprise partners. Public paid API access is coming. For developers, this is a blocker for now, you can use Muse Spark through the Meta AI app, but cannot integrate it into your own products yet.

Who Should Use Each Model?

The Meta Strategy Behind Muse Spark

This is not just a model launch, it is Meta's attempt to rebuild AI credibility after Llama 4 Maverick disappointed benchmarks. The $14B Alexandr Wang bet, the creation of Meta Superintelligence Labs, the shift from open-source to proprietary: these signal that Zuckerberg treats AI as existential to Meta's business.

Free access to Muse Spark drives Meta AI adoption across 3.27 billion daily active users across Facebook, Instagram, WhatsApp, and Messenger. Every interaction trains better models. Every Meta AI session generates data. The free model is a data acquisition strategy as much as a product launch.

Wall Street responded positively. Analysts noted the health benchmark leadership as a potential wedge into the healthcare AI market, a $45B+ opportunity where no single company has established dominant position.

Bottom Line

Muse Spark is not the best AI model. But it is the best free AI model, and for health applications specifically, it is the best model period, paid or otherwise.

GPT-5.4 Thinking remains the overall frontier leader. Gemini 3.1 Pro ties it on the Intelligence Index and beats Muse Spark on coding and reasoning. Claude Sonnet 4.6 leads on agentic tasks alongside GPT-5.4.

Use Muse Spark when you need strong health AI, high token efficiency, or cannot afford a $200/month Pro subscription. It is a real addition to the frontier, just not the top of it. Yet.