Meta Muse Spark: The AI Redefining Multimodal Reasoning

What is Muse Spark?

At the heart of Muse Spark lies Meta's latest language model, the first of its new Muse family. Meta presents Muse Spark as a model that is both small and fast, capable of handling more complex reasoning tasks. This means it is not just another chatbot brain. It is positioned as the foundational layer of a smarter Meta AI, capable of tackling difficult questions, understanding images, and supporting more complex tasks within the Meta ecosystem.

This clearly distinguishes Muse Spark. Meta is not introducing it as a mere lab demonstration intended to impress AI researchers online. Muse Spark is presented as a product-oriented model that is already powering the Meta AI application and website. The company also claims that the model is designed for multimodal tasks, stronger reasoning, and faster responses, with larger Muse models already in development. In other words, Muse Spark is Meta's attempt to create an AI model that genuinely assists users in the applications they use daily.

Muse Spark: Features

Meta has kept the overall feature set of Muse Spark relatively focused at launch. Instead of offering a long list of impressive capabilities, it highlights three major areas where the model is expected to be useful.

Contemplation Mode

One of Muse Spark's standout features, Contemplation Mode, orchestrates multiple agents reasoning in parallel. Meta claims this allows the model to tackle more challenging tasks with deeper reasoning. The company positions it as a way for Muse Spark to compete with the advanced reasoning modes of leading models like Gemini Deep Think and GPT Pro.

Meta also backs this claim with numbers, indicating that Contemplation Mode achieves 58% on the final human exam and 38% in cutting-edge scientific research.

Muse Spark is also designed to work with visual information from the outset. Meta states that the model can handle visual STEM questions, entity recognition, and localization, making it useful across a broader range of tasks than text-only systems. This capability also fuels more interactive use cases, such as creating mini-games or helping users troubleshoot household devices with dynamic annotations.

Health

This is a new focus and one of the key areas that Meta has clearly prioritized. The company claims to have collaborated with over 1,000 doctors to develop training data that enhances Muse Spark's health reasoning capabilities. As a result, the model is designed to provide more factual and comprehensive health-related answers. Meta also asserts that Muse Spark can generate interactive displays to explain elements like the nutritional content of foods or the muscles activated during exercise.

Overall, these features clearly illustrate the direction Meta is taking with Muse Spark. This model is positioned as a more thoughtful, visual, and practical system for everyday life.

Muse Spark: Architecture

Meta explains Muse Spark through three scaling axes: pre-training, reinforcement learning, and reasoning at test time. In other words, this is how the company demonstrates where the model's core intelligence comes from. It also indicates how this intelligence is enhanced after the initial training and how it becomes more efficient when responding to user queries.

At this stage, Muse Spark builds its core capabilities in multimodal understanding, reasoning, and coding. Meta states it has rebuilt the entire architecture over the past nine months, improving the optimization process and data curation. According to the company, these changes allow Muse Spark to achieve the same level of capability with much less computational power than Llama 4 Maverick. This is a significant claim, as it suggests that Muse Spark is not only more powerful but also much more efficient.

Reinforcement Learning

After pre-training, Meta uses reinforcement learning to further enhance the model. The company claims that this phase offers smooth and predictable gains, despite the fact that large-scale reinforcement learning is often unstable. More importantly, Meta asserts that these gains are not limited to training data alone. Muse Spark also improves on evaluation tasks set aside. This suggests that the additional training generalizes beyond the exact problems it has already encountered.

Reasoning at Test Time

This is the part that controls how Muse Spark "thinks" before responding. Meta indicates that it uses thinking time penalties to ensure that the model uses its reasoning tokens more efficiently, rather than simply producing long chains of thought. The company also employs multi-agent orchestration here, allowing multiple parallel agents to work together on a difficult problem. According to Meta, this gives Muse Spark better performance with comparable latency. This will be very useful if the company wishes to offer this capability to billions of users.

Muse Spark: Performance on Benchmarks

Muse Spark appears strongest in the areas that Meta emphasizes the most: multimodal understanding, health, and deeper reasoning through Contemplation Mode. The model scores 78.4 on MedXpertQA (MM), supporting Meta's claim that health is one of the model's key domains. Its Contemplation Mode bolsters the reasoning narrative, pushing Muse Spark to 50.2 on the final human exam (without tools) and 38.3 on cutting-edge scientific research, outpacing some of the top competitors in these comparisons.

However, it is worth noting that the results do not show a clean sweep of the benchmarks. In some broader reasoning tests, coding, and agentic evaluations, stronger rivals remain ahead, particularly on tests like ARC AGI 2 and certain aspects of coding performance. Thus, the conclusion is quite clear: Muse Spark does not yet seem to be the strongest leading model overall. However, it shows clear and credible strength in the specific areas for which Meta appears to have designed it.

Muse Spark: How to Access It

Meta's new AI model is already accessible. You can access it in several ways:

• Go to the meta.ai platform and use it via the chat interface
• Download the Meta AI app on your phone and use it

Meta has also announced that it is opening a private preview of the API to selected users, meaning broader developer access is still limited for now.

Let's Try Muse Spark

Once you access Muse Spark, you will realize its true beauty. It brings back the traditional AI chatbot interface in a clean and minimalist way, without unnecessary options or tools to choose from. Just 2 modes – Create or add media/files to your chat. That's it!

With this simplicity and its claims in mind, we put Muse Spark through a series of tests to verify its capabilities. Here’s how it performed.

Task 1: Text Extraction from an Image

"Extract all the text from this image and formulate a WhatsApp message to send in groups using the information."

Muse Spark handled the text extraction task competently and with good accuracy. The model succeeded...