US-based tech giant Meta has launched Muse spark. This is the first model in its new Muse family developed by Meta Superintelligence Labs. The new launch is positioned as a step toward building personal superintelligence. Muse Spark is a natively multimodal reasoning model with support for tool use, visual chain-of-thought reasoning, and multi-agent orchestration. The model is now available through Meta AI and the Meta AI app with a private API preview for select users.
The tech giant claims that the Muse Spark is a part of a broader overhaul of its AI stack, covering model training, infrastructure, and deployment. The company is further investing in large-scale systems, for example, the Hyperion data center to support future models in the Muse family.
Today we’re introducing Muse Spark, our most powerful model yet, giving you a faster and smarter Meta AI.
Muse Spark currently powers the Meta AI app and website and will be rolling out to @whatsapp, @Instagram, @facebook, @messenger, and AI glasses in the coming weeks.…
— Meta Newsroom (@MetaNewsroom) April 8, 2026
Muse Spark: Meta’s New Step Toward Superintelligence
Muse Spark is designed to process and reason across text, images, and tools. It integrates multimodal understanding with agent-based workflows, enabling it to solve complex problems using multiple reasoning agents working in parallel.
The company has also rolled out a “Contemplating mode” which allows multiple agents to reason simultaneously. This improves performance in complex tasks without significantly increasing latency. As per the evaluation done by the company itself, it achieved 58 per cent on Humanity’s Last Exam and 38 per cent on Frontier Science Research.
The model is designed for real-world applications such as analysing objects, creating generative interactive tutorials, and assisting with health-related queries. The company claims that it has worked with more than 1,000 physicians to enhance the model’s ability to provide structured and factual health insights.
Multimodal Reasoning and Contemplating Mode
The company stated that Muse Spark delivers strong results across multimodal and reasoning benchmarks, outperforming competing models such as Gemini 3.1 Pro, GPT-5.4, and Grok 4.2 in several tests.
In CharXiv reasoning, the model scored 86.4 whereas ChatGPT 5.4 scored 82.8, Gemini 3.1 Pro scored 80.2, and Grok 4.2 scored 60.9. In SimpleVQA it scored 71.3 while in ZeroBench it scored 33.0 which was more than Gemini’s 29.0 but trailing GPT-5.4’s which scored 41.0
In terms of health-related benchmarks, Muse Spark recorded 42.8 in HealthBench Hard, which is higher than Gemini 3.1 Pro’s 20.6 and Grok 4.2’s 20.3 score and GPT-5.4’s 40.1 mark.
However, the company’s own data shows mixed results in some areas. In MMMU Pro and ARC AGI 2 Gemini and GPT models scored higher scores than Muse Spark, indicating areas where Muse Spark still lags.
Benchmark Performance and Competition
Muse Spark is developed on a redesigned AI stack that enhances efficiency across pretraining, reinforcement learning, and test-time reasoning. Meta claims it can achieve similar capability levels with significantly lower computing compared to earlier systems such as Llama 4 Maverick.
Meta highlights improvements in reinforcement learning stability and test-time reasoning efficiency. Multi-agent orchestration is used to scale reasoning performance without exceeding response time.
Safety, Testing, and Availability
The company claims that newly launched Muse Spark has undergone safety evaluation under its Advanced AI Scaling Framework. The model demonstrates refusal behaviour in high-risk domains and does not show autonomous capabilities linked to harmful scenarios.
The model is available via Meta AI platforms with features such as Contemplating mode rolling out gradually.
Syed Ziyauddin is a media and international relations enthusiast with a strong academic and professional foundation. He holds a Bachelor’s degree in Mass Media from Jamia Millia Islamia and a Master’s in International Relations (West Asia) from the same institution.
He has work with organizations like ANN Media, TV9 Bharatvarsh, NDTV and Centre for Discourse, Fusion, and Analysis (CDFA) his core interest includes Tech, Auto and global affairs.
Tweets @ZiyaIbnHameed
