Meta has quietly pivoted its AI strategy away from open-source dominance toward proprietary models that prioritize efficiency over brute-force computing power. The company's new Muse Spark model demonstrates that the future of advanced reasoning may not require the massive computational budgets everyone assumed necessary. By using a technique called "thought compression," Muse Spark achieves reasoning capabilities comparable to industry leaders while consuming over an order of magnitude less compute than Meta's previous flagship, Llama 4 Maverick .

This shift represents a fundamental rethinking of how artificial intelligence companies approach the inference phase, the moment when a model actually processes your question and generates an answer. Rather than throwing more computing power at the problem, Meta's new approach penalizes models during training for excessive "thinking time," forcing them to solve complex problems with fewer reasoning tokens without sacrificing accuracy .

What Is Test-Time Compute and Why Does It Matter?

Test-time compute refers to the computational resources a model uses when answering your question, as opposed to the resources needed to train the model initially. For years, the AI industry focused almost exclusively on training efficiency and model size. But as reasoning models like OpenAI's o-series and Google's Gemini Deep Think gained prominence, companies realized that the real bottleneck wasn't training; it was inference scaling, the ability to allocate more computing power at the moment of use to tackle harder problems .

Meta's approach inverts this logic. Instead of asking "how much compute can we throw at inference," the company asked "how can we make inference smarter with less compute?" The answer was thought compression, a training technique that rewards models for reaching correct answers efficiently. This matters because it directly impacts the cost of deploying AI systems at scale. If you can achieve the same reasoning quality with 10 times less computing power, you've just reduced operational costs by roughly 90 percent .

How Does Muse Spark Achieve Superior Efficiency?

• Thought Compression: During reinforcement learning, the model is penalized for excessive "thinking time," forcing it to solve complex problems with fewer reasoning tokens while maintaining accuracy levels comparable to less efficient competitors.
• Visual Chain of Thought: Unlike previous models that stitched vision and text together, Muse Spark was rebuilt from the ground up to integrate visual information across its internal logic, enabling the model to annotate dynamic environments and reason through spatial problems.
• Parallel Multi-Agent Reasoning: A new "Contemplating" mode orchestrates multiple sub-agents to reason in parallel, allowing Meta to compete with extreme reasoning models without proportionally increasing computational demands.

The practical result is striking. Muse Spark achieved a score of 42.8 on the Humanity's Last Exam benchmark, a multidisciplinary evaluation designed to test reasoning across diverse domains, using substantially less compute than competitors . In figure understanding tasks, the model scored 86.4, significantly outperforming Claude Opus 4.6 at 65.3 and matching or exceeding other industry leaders .

Where Does Muse Spark Stand Against the Competition?

According to independent auditing from Artificial Analysis, Muse Spark achieved an overall Intelligence Index score of 52, placing it within striking distance of the industry's most elite systems . For context, Meta's previous flagship, Llama 4 Maverick, debuted in 2025 with an index score of just 18. By nearly tripling its performance, Muse Spark now sits behind only Gemini 3.1 Pro Preview at 57, GPT-5.4 at 57, and Claude Opus 4.6 at 53 .

The model shows particular strength in specialized domains. In health-related benchmarks, Muse Spark achieved 42.8 on HealthBench Hard, a massive lead over Claude Opus 4.6 at 14.8 and Gemini 3.1 Pro at 20.6, likely a result of Meta's collaboration with over 1,000 physicians . On multimodal medical questions, it scored 78.4, comfortably ahead of Opus 4.6 at 64.8 and Grok 4.2 at 65.8 .

However, Muse Spark shows weakness in abstract reasoning. On the ARC AGI 2 benchmark, which tests abstract reasoning puzzles, the model scored 42.5, far behind Gemini 3.1 Pro at 76.5 and GPT-5.4 at 76.1 . This suggests that while the model excels at reasoning through visual and domain-specific problems, it struggles with the kind of novel problem-solving that doesn't fit neatly into existing categories.

What Does This Mean for the Future of AI Inference?

Muse Spark's efficiency breakthrough signals a potential inflection point in how the AI industry approaches scaling. For the past two years, the dominant narrative has been that bigger models with more reasoning time equals better performance. OpenAI's o-series models and Google's Deep Think variants both rely on allocating substantial compute at inference time to achieve superior reasoning .

Meta's approach suggests an alternative path: smarter allocation of compute rather than more compute. If thought compression becomes a standard training technique across the industry, it could reshape the economics of AI deployment. Companies running large-scale AI systems would face a choice: invest in more expensive hardware to support longer reasoning chains, or invest in better training techniques that achieve similar results with less inference compute.

The launch of Muse Spark also marks a strategic departure for Meta. The company built its reputation in the AI era on open-source models like Llama, which democratized access to advanced AI capabilities. Muse Spark is proprietary, available only through Meta's AI app, website, and a private API preview to select users . This shift suggests that Meta's leadership, under newly appointed Chief AI Officer Alexandr Wang, believes the company's competitive advantage lies not in open-source community building but in proprietary models optimized for specific use cases like healthcare and personal superintelligence .

"Muse Spark is the most powerful model that Meta has released," Wang stated, noting it has "support for tool-use, visual chain of thought, and multi-agent orchestration."

Alexandr Wang, Chief AI Officer at Meta

The broader implication is that test-time compute scaling, once seen as the primary lever for improving AI reasoning, may be reaching diminishing returns. As models become more sophisticated at allocating their reasoning resources efficiently, the competitive advantage shifts from raw computational power to algorithmic innovation. This could level the playing field for companies without access to the largest computational budgets, allowing them to compete through smarter training techniques rather than deeper pockets .