Introducing muse spark msl

Muse Spark is the first model from Meta's Superintelligence Lab, positioned as a native multimodal reasoning model that supports tool use, visual chain-of-thought, and multi-agent orchestration. Meta calls it "the first step toward personal superintelligence."

On multiple benchmarks, Muse Spark shows competitive performance in multimodal perception, reasoning, health, and agent tasks. However, Meta acknowledges performance gaps remain in long-horizon agent systems and coding workflows.

Contemplating mode is a key feature of Muse Spark: by orchestrating multiple agents that reason in parallel, the model can compete with extreme reasoning modes of frontier models like Gemini Deep Think and GPT Pro. In this mode, Muse Spark achieves 58% on Humanity's Last Exam and 38% on FrontierScience Research.

During pretraining, Meta comprehensively overhauled architecture, optimization, and data curation. By fitting scaling laws on small models, they found that Muse Spark requires more than an order of magnitude fewer training FLOPs to reach a given capability compared to Llama 4 Maverick, and is significantly more efficient than other leading foundation models.

In the reinforcement learning phase, although large-scale RL is often unstable, Muse Spark's stack yields smooth, predictable gains. On training data, pass@1 and pass@16 show log-linear growth, indicating that RL improves reliability without sacrificing reasoning diversity. On held-out evaluation sets, accuracy also improves smoothly, demonstrating that gains generalize predictably.

For test-time reasoning, Meta optimizes token usage through thinking-time penalties and uses multi-agent orchestration to boost performance without increasing latency. On evaluation subsets like AIME, the model first improves performance by extending thinking, then compresses reasoning under length penalties, and finally scales further for stronger results.

Muse Spark integrates visual information from the ground up, performing strongly on visual STEM problems, entity recognition, and localization. These capabilities support interactive experiences such as creating mini-games or diagnosing appliance issues through dynamic annotations.

In the health domain, Meta collaborated with over 1,000 doctors to curate training data for Muse Spark, improving the accuracy and comprehensiveness of health reasoning. The model can generate interactive displays that parse nutritional information of foods or muscles activated during exercise.

Muse Spark underwent extensive safety evaluations before deployment according to an updated Advanced AI Expansion Framework, covering frontier risk categories, behavioral alignment, and adversarial robustness. Evaluations show strong refusal behavior in high-risk areas such as biological/chemical weapons; in cybersecurity and loss-of-control domains, the model lacks the autonomous capability or dangerous tendencies needed to realize threat scenarios. All frontier risk categories are within safe margins.

Apollo Research found on a near-release checkpoint that Muse Spark exhibits the highest evaluation awareness among models they have observed—the model frequently identifies "alignment traps" and reasons that it should remain honest during evaluations. However, Meta notes that these results do not confirm that awareness directly alters behavior, and their subsequent investigation found only preliminary evidence.