Meta has just overhauled its Meta AI app visually and under the hood. With this update, Meta has started using a completely new LLM that is closed sourced. Previously called Avocado, the new model is now called Muse Spark and will make up a larger family of models called Muse. The new model was developed by Meta Superintelligence Labs and is natively multimodal and reasoning. It also supports tool-use, visual chain of thought, and multi-agent orchestration.
The new Muse Spark model is now available to all on meta.ai and the Meta AI app, just ensure you’ve updated it from the app store. It is also providing a private API preview for select users
Meta said that Muse Spark is just the first step on its “scaling ladder” and the first product it has built as part of its ground-up overhaul of its AI efforts. To help scale up further, the company is making investments across the entire stack, from research and model training up to infrastructure, including its Hyperion data center.
Muse Spark (Thinking) offers competitive results when compared to Opus 4.6 Max, Gemini 3.1 Pro High, GPT 5.4 Xhigh, and Grok 4.2 Reasoning. Meta says that it is very capable at multimodal perception, reasoning, health, and agentic tasks, however, acknowledged performance gaps in long-horizon agentic system and coding workflows. It is investing to try to close these gaps, however.
In addition to Instant and Thinking modes, Meta is working on another mode, which it is still to release, called Contemplating mode. This mode orchestrates multiple agents that reason in parallel allowing Muse Spark to compete with extreme reasoning modes of frontier models such as Gemini Deep Think and GPT Pro. It brings big improvements on challenging benchmarks scoring 58% in Humanity’s Last Exam and 38% in FrontierScience Research.
If you want to take a deeper dive into this interesting work from Meta, then check out the company’s blog post.