Meta Superintelligence Labs unveils Muse Spark, its first frontier model. Surprise: it is neither a competitor to ChatGPT nor a model for developers.
Millions of dollars in salaries to recruit the best researchers and billions of dollars in investments to build an AI infrastructure, all to launch the Meta Superintelligence Labs (MSL) laboratory in 2025. This riot of resources employed by Meta has finally delivered a significant result: Muse Spark, its first frontier model. In the benchmarks, the model is generally behind Gemini 3.1 Pro but at a level close to, or even higher than, that of Anthropic’s Opus 4.6. Surroundingly, with this launch, Meta is making a 180-degree turn in its generative AI strategy: no more open source, Muse Spark will be proprietary. Decryption.
A model designed for science and health…
On pure reasoning and code, Muse Spark clearly remains behind. Gemini 3.1 Pro dominates GPQA Diamond (94.3% vs. 89.5%), LiveCodeBench Pro (82.9% vs. 80%) and agentic code benchmarks like Terminal-Bench 2.0 (68.5% vs. 59%). GPT 5.4 and Opus 4.6 also do better on other benchmarks.
Where Muse Spark really differentiates itself is on health and visual perception. The model takes the lead on HealthBench Hard (open medical answers: 42.8%, far ahead of Gemini 3.1 at 20.6%), MedXpertQA multimodal (medical multiple choice questions with images) and CharXiv Reasoning (comprehension of scientific figures).
To perform on scientific reasoningMeta has developed a “contemplating” mode capable of orchestrating several agents in parallel to maximize the intelligence of the model. The latter allows Meta’s AI to beat Gemini Deep Think and GPT 5.4 Pro on Humanity’s Last Exam (expert-level multidisciplinary reasoning) and FrontierScience Research (cutting-edge scientific research questions).
…and connected health objects
With more than 1,000 doctors mobilized to curation of data from the model training dataset, Meta is truly focusing the development of its AI on health. Muse Spark is probably one of the best models on the market in the medical sector. And this is no coincidence: in all its communication, Meta hammers home the concept of “personal superintelligence”, an AI that understands your immediate environment, analyzes your health data and supports you on a daily basis. Marked by the wave of autonomous agents (OpenClaw & co) and with the acquisition of Manus, Meta’s strategy seems clear: to become the platform for the personal agents of tomorrow.
A clearly B2C orientation very consistent with the Meta ecosystem. The partnership with Ray-Ban, which already embeds Meta AI in its connected glasses, takes on a whole new dimension if the underlying model excels in visual perception and health. We can imagine glasses capable of analyzing a meal tray, tracking physical activity or providing contextual recommendations in real time. A positioning (soon?) in healthtech hardware, whether through a partnership or internal development, could be the real reason for Muse Spark.
Furthermore, the fact that Meta has gone all out on the security of the model confirms this B2C orientation: the model displays the strictest safeguards on the market. A level of caution which is undoubtedly explained by access to the gigantic dataset from Meta applications, but above all by the need to massively secure a model intended to be deployed to billions of general public users (via its applications). The model is not designed for business. The API, first opened in private preview, will probably be used, above all, to distribute this personal intelligence to its partners.
After the flop of the metaverse, which swallowed up tens of billions of dollars, the path this time seems more promising. Meta is investing in an area where the benefits will be concrete and where the demand is real. However, we will have to maintain agility: Google, positioned in the same segment of personal intelligence with Gemini and its own ecosystem, also has the right data and will not let this happen. The race for the personal assistant has only just begun.
