Meta has released a new model. This time it's called Muse Spark, positioned as' personal super intelligence '.
To be honest, the term 'personal super intelligence' sounds a bit marketing oriented, but after watching the entire press conference, I think there are two points worth mentioning: one is computational efficiency, and the other is healthy reasoning.
First, let's talk about the hard indicator of computing power efficiency
Meta's own comparative data: To achieve the same level of performance, Muse Spark requires only 1/10 of the computing power of Llama4 Maverick.
If this number is true, then it is really exaggerated. The cost of computing power is one of the core bottlenecks in the application of large models. Being able to save 90% of computing power means:
Individual users are more likely to run locally
Significant reduction in deployment costs for enterprises
Ability to implement intelligent applications on more devices
This may be Meta's technological confidence in positioning itself as "personal super intelligence".
What is the Contemplating Mode?
Muse Spark has a special mode called Contemplating, which uses a multi-agent parallel inference architecture.
Benchmark test score:
Humanity's Last Exam：58%
FrontierScience Research：38%
Meta directly benchmarks Gemini 3.1 Deep Think and GPT 5.4 Pro. This benchmark selection is quite interesting, as they are all models that focus on deep reasoning.
The idea of multi-agent parallel reasoning is essentially to have multiple "thinkers" work simultaneously and then synthesize the results. In theory, it can deal with more complex problems than the serial reasoning of a single model.
Taking photos to generate Sudoku, this demonstration is a bit interesting
Muse Spark adopts a native multimodal architecture - visual information is integrated from the bottom, not post stitching.
Presentation at the press conference: The user takes a photo and the model automatically generates a complete Sudoku game.
This demonstration may seem simple, but it actually demonstrates two abilities:
Visual understanding: able to recognize the content in photos
Logical generation: able to generate structured games based on understanding
This is closer to practical application scenarios than those demonstrations that rely on pictures to speak.
Joint training with over 1000 doctors
This is the most interesting part in my opinion: Muse Spark has conducted specialized training in the field of health, collaborating with over 1000 doctors.
Specific abilities:
Users upload food photos or data, and models analyze nutritional components
Mark recommended and non recommended foods with red and green dots
Generate highly interactive displays of health information
This direction was chosen very cleverly. Health is one of the most sensitive yet valuable areas for AI applications. Collaborating with professional doctors for training not only enhances professionalism but also reduces risks.
However, the statement 'health consultants become professional doctors in seconds' is still a bit exaggerated. AI can assist in health decision-making, but there is still a long way to go to replace professional diagnosis.
Already online, API preview synchronized startup
Muse Spark has been launched on the meta.ai and Meta AI mobile applications. Simultaneously open private API preview to some users.
The Contemplating mode will gradually be pushed to more users. Meta stated that future Muse series products will continue to iterate around "personal super intelligence".
What is Meta's abacus?
From the positioning and functionality of Muse Spark, Meta's strategy is very clear:
Differentiated competition: not comparing parameter scale with large models, comparing computational efficiency and practical applications
Scenario entry: Vertical fields such as health and education are easier to implement than general abilities
Ecological binding: Keeping users in their own ecosystem through meta.ai and Meta AI applications
This approach is more pragmatic than simply focusing on the size of the parameters.
