For the last few years, Meta’s approach to artificial intelligence was defined entirely by its open-source Llama models. While they were incredibly popular among developers, the company always seemed to be playing catch-up with the absolute top-tier proprietary models from OpenAI and Google. That dynamic is officially changing. Meta has completely retooled its artificial intelligence division, launching the new Meta Superintelligence Labs, and their first major release is a brand-new, natively multimodal large language model called Muse Spark.
This marks a significant, deliberate pivot for the social media giant. Muse Spark is not just another iterative update; it represents a ground-up overhaul of the company’s entire AI stack. More importantly, unlike the widely accessible Llama series, this new model is strictly proprietary, placing Meta in direct, head-to-head competition with heavy hitters like GPT-5.4 and the latest Gemini 3.1 Pro.
A ‘People-First’ Multimodal Approach
While other tech giants are racing to build enterprise-grade agents, Meta claims Muse Spark is explicitly designed around a “people-first” philosophy, aiming to function as a highly personal superintelligence. Because it is a natively multimodal system, it does not just process text; it can simultaneously understand complex visual inputs, spatial context, and real-world scenes.
In practical terms, this allows the AI to step out of a standard chat window and actively assist with daily life. For instance, users can snap a photo of a broken home appliance and ask the model to troubleshoot the issue using dynamic, interactive annotations. Meta has also heavily focused on making the model a reliable health and wellness assistant. By collaborating directly with over a thousand physicians during the training process, Muse Spark is uniquely positioned to analyse physical activity or the nutritional content of a meal, generating highly factual, structured insights based on your specific context.
Deep Reasoning with Contemplating Mode
To compete at the absolute frontier of AI performance, a model needs to be able to pause and think through complex, multi-step problems. To achieve this, Meta is introducing a dedicated Contemplating mode alongside its standard, high-speed Instant mode.
When faced with a particularly difficult prompt, Contemplating mode effectively launches multiple subagents that reason in parallel to tackle the issue. This allows the model to act almost like a project manager, breaking down a complicated request—like drafting a comprehensive travel itinerary while comparing different cities and locating kid-friendly activities—into manageable tasks before delivering a cohesive final answer. According to Meta, this multi-agent orchestration allows Muse Spark to rival the extreme reasoning capabilities found in specialised models like Gemini Deep Think.
If you want to test out this new digital assistant, you will not have to wait. Muse Spark is officially rolling out today across the United States within the Meta AI application and through the meta.ai website, with the advanced Contemplating mode scheduled for a gradual release over the coming weeks.
