Last April Meta When Llama 4 was launched, it was nothing short of a disaster. At the time, it was exposed that the benchmark scores were fabricated, and the most powerful Behemoth version kept getting delayed repeatedly, which caused CEO Mark Zuckerberg to nearly completely lose confidence in the internal team. Meta then made a major move by hiring the former Scale AI CEO Alexandr Wang established a brand new Meta Superintelligence Labs division, ripped out and rebuilt the entire AI tech stack, and spent a full 9 months building new models, new infrastructure, and new data processing pipelines from scratch.
After nine months of development, the results were finally unveiled earlier today when Meta officially launched the all-new Muse Spark model.
Meta Releases New AI Model Muse Spark: Leads Competitors in Health and Science Reasoning, But Still Lags in Coding
Muse Spark This is the first model launched by Meta Superintelligence Labs, and unlike previous Llama series, Meta is no longer just chasing size. Instead, it has adopted a more systematic progressive scaling strategy—first releasing a small but refined model to validate technical feasibility, then gradually scaling up.
In simple terms, Muse Spark is a native multimodal reasoning model that understands both text and images. It offers two modes: “Instant” and “Thinking”:
Additionally, there’s a particularly special “Contemplating” mode, Meta’s exclusive design that simultaneously activates multiple AI agents to collaborate on tackling more difficult problems, similar to Gemini 3.1’s Deep Think and GPT 5.4 Pro.
Like when planning a Florida family vacation, one agent handles the itinerary, another compares Orlando with the Florida Keys, and a third finds kid-friendly activities—all running simultaneously to give you faster, more complete answers.
Meta also specifically emphasized Muse Spark’s capabilities in the health domain, collaborating with over 1,000 doctors to help organize specialized training data, enabling Muse Spark to respond to health-related questions, including analyzing charts and images.
One more point worth mentioning is that, according to Meta’s official statement, Muse Spark requires more than 10x less computational resources to achieve comparable performance compared to the previous generation Llama 4 Maverick, representing a significant breakthrough in training efficiency.
In terms of benchmark data, according to Artificial Analysis’s AI Intelligence Index, Muse Spark scored 52 points, ranking within the top 5 globally, but still trails GPT-5.4 and Gemini 3.1 Pro (both at 57 points), and slightly trails Claude Opus 4.6 (53 points), while outperforming Claude Sonnet 4.6 and other major competitors:
Meta has shared several Muse Spark test data points. Here are the key highlights:
- In the health domain HealthBench Hard test, Muse Spark scored 42.8 points, surpassing GPT-5.4’s 40.1 points
- Chart comprehension: It also scored 86.4 to take first place on the CharXiv Reasoning test, beating GPT-5.4’s 82.8 and Gemini 3.1 Pro’s 80.2.
- In the AI search domain, during DeepSearchQA testing, Muse Spark also achieved the highest score of 74.8, slightly outperforming Opus 4.6’s 73.7 and GPT-5.4’s 73.6.
In Contemplating mode, the scientific reasoning Humanity’s Last Exam test scored 50.2%, and with tools it reached 58%, both outperforming Gemini 3.1 Deep Think and GPT 5.4 Pro:
However, Muse Spark is noticeably weaker at coding, scoring only 59.0 on the Terminal-Bench 2.0 test, compared to GPT-5.4’s 75.1 and Gemini 3.1 Pro’s 68.5. It’s even further behind on abstract reasoning with ARC-AGI-2, where Muse Spark scored 42.5 while GPT-5.4 and Gemini 3.1 Pro scored 76.1 and 76.5 respectively.
Overall, Muse Spark is quite strong in health, chart comprehension, and AI search, and its multimodal capabilities are decent. However, in areas like coding, abstract reasoning, and agentic tasks, there’s still a considerable gap compared to top competitors.
Meta has itself acknowledged that there’s still room for improvement in “long-running agentic tasks” and “code workflows.”
Muse Spark is now live, available through the Meta AI app and meta.ai website, offering both Instant and Thinking modes, with the initial rollout in the United States.
Over the next few weeks, Muse Spark will also be gradually rolled out to other platforms under its umbrella, including WhatsApp, Instagram, Facebook, Messenger, and Meta’s AI glasses. A private API preview will also be made available to select partners, though there’s no public API pricing information available yet.
One thing to note is that unlike previous Llama series, Muse Spark is closed-source this time and will not release the model’s weights and code.
Source: KOCPC Chinese