Meta Platforms Inc. today debuted a new reasoning model, Muse Spark, that is highly adept at answering health questions and analyzing multimodal data. 

The company will roll out the algorithm to its consumer-focused Meta AI artificial intelligence service over the next few weeks. In addition, Meta is making Muse Spark available through developers through an application programming interface. The API is in private preview.

Meta’s progress in AI, following a series of stumbles, encouraged investors. Its stock rose 6.5% today, though part of that was the result of a huge day for the overall market, with the tech-heavy Nasdaq up 2.8% as the Iran war’s impact was seen as easing at least temporarily.

The company says that Muse Spark outperforms Claude 4.6 Opus, Gemini 3.1 Pro and GPT 5.4 across several benchmarks. One of them is HealthBench Hard, an evaluation that measures artificial intelligence models’ ability to answer medical questions. Muse Spark beat the score of the runner-up, GPT 5.4, by more than 2%.

The model’s performance is partly the fruit of a clinical training dataset that Meta compiled with the help of over 1,000 physicians. The dataset was developed as part of a broad revamp of the company’s AI development workflow. According to the Facebook parent, its engineers also enhanced its model architecture and post-training workflow.

“We can reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick,” Meta stated in a blog post today. “This improvement also makes Muse Spark significantly more efficient than the leading base models available for comparison.”

According to Meta, scientific chart analysis is another task that Muse Spark performs better than the competition. It bested Opus 4.6 and other rivals on CharXiv Reasoning, a benchmark dataset that comprises technical graphs. That visual reasoning capability carries over well to other use cases. Users of the Meta AI app can upload a photo of a grocery store shelf and ask it to estimate the calorie count of each food item.

Meta also tested Muse Spark across more than a half-dozen other benchmarks. It came within a few percentage points of Opus 4.6, Gemini 3.1 Pro and GPT 5.4 in many cases. There were multiple evaluations in which Muse Spark outperformed at least one of the competing models. The benchmarks covered use cases such as code generation, robot navigation and tool use.

Muse Spark can boost its output quality by activating a setting called Contemplating mode. The feature launches multiple AI agents that break down a task into substeps and carry them out in parallel. Meta says that the technology increased Muse Spark’s score on HLE, one of the AI ecosystem’s most difficult benchmarks, by about 8%.

Muse Spark is the first in a planned series of multimodal reasoning models. “We’re on a predictable and efficient scaling trajectory,” Meta stated in today’s blog post. “We look forward to sharing increasingly capable models on the path to personal superintelligence soon.”

Image: Meta