Meta Launches Muse Spark, Its Most Capable AI Yet—But Gemini 3.1 Pro Still Leads the Pack

Meta released Muse Glow on Wednesday, marking the very first design developed by Meta Superintelligence Labs– the group put together 9 months earlier under Chief AI Officer Alexandr Wang after Meta’s $14 billion Scale AI acquisition. It’s live now at meta.ai and the Meta AI app, with a rollout to Facebook, Instagram, and WhatsApp can be found in the next couple of weeks.

This isn’t simply another chatbot upgrade or a brand-new variation of Llama. Muse Glow is natively multimodal– it processes images, text, and voice from the ground up, instead of bolting vision onto an existing text design. It includes visual chain-of-thought, tool-use assistance, and something Meta is calling “Considering mode”: a setup that runs several AI representatives in parallel to take on more difficult issues. That’s Meta’s response to the prolonged thinking modes from Google’s Gemini Deep Believe and OpenAI’s GPT Pro.

” Muse Glow is the initial step on our scaling ladder and the very first item of a ground-up overhaul of our AI efforts,” Meta composed in a main statement. “To support more scaling, we are making tactical financial investments throughout the whole stack– from research study and design training to facilities, consisting of the Hyperion information center.”

The business dealt with more than 1,000 doctors to curate training information for Muse Glow’s medical thinking. The outcomes on HealthBench Hard– an open-ended health questions standard– stand out: Muse Glow scored 42.8, compared to 40.1 for GPT 5.4 and simply 20.6 for Gemini 3.1 Pro. That’s not a limited distinction.

On agentic search (DeepSearchQA), Muse Glow likewise leads with 74.8, beating Gemini (69.7) and GPT 5.4 (73.6 ). On CharXiv Thinking– figure understanding from clinical documents– it scored 86.4, the greatest throughout the designs in the contrast.

For those into jailbreaking AI, the design was broken open within minutes:

However excellent is n’t the like fantastic. The total benchmark photo reveals Gemini 3.1 Pro still running ahead on a lot of classifications. The space is most noticeable on ARC AGI 2, the abstract thinking puzzle standard: Gemini scored 76.5 to Muse Glow’s 42.5. (* )On coding( LiveCodeBench Pro), Gemini’s 82.9 surpasses Meta’s 80.0. On MMMU Pro– multimodal understanding– Gemini scored 83.9 versus 80.4. Meta’s own blog site acknowledges existing efficiency spaces in long-horizon agentic systems and coding workflows.

There’s likewise a noteworthy tactical shift baked into this launch. Muse Glow is a closed design– its architecture and weights will not be revealed. That’s a sharp departure from Llama, which developed Meta’s credibility in open AI circles. After Llama 4’s underwhelming reception previously this year, Meta appears to have actually chosen the next chapter requires to be composed in a different way.

The business states it wants to open-source future variations of Muse, however for now the code remains inside Meta. The tech giant’s stock climbed up almost 9% on Wednesday following the statement, and completed the trading day up 6.5% to a cost of$ 612.42.

“Considering mode” utilizes parallel representative orchestration to press the design’s ceiling greater. Because setup, Muse Glow struck 58% on Mankind’s Last Examination and 38 %on FrontierScience Research study– area that makes it competitive with the most capable variations of Gemini and GPT, instead of their basic releases.

Meta is likewise presenting a shopping assistant that compares items and links straight to purchases, and prepares to bring Muse Glow to Facebook, Instagram, and WhatsApp in the coming weeks– following the exact same script carried out because Llama 3, putting it in front of more than 3.5 billion users. A personal API sneak peek is opening to pick designers.

The design was integrated in 9 months, internally codenamed Avocado, with Meta declaring that its brand-new pretraining stack can reach the exact same ability level as Llama 4 Radical utilizing over 10 times less calculate.

Muse Glow is explained internally as a “little and quick “initial step in the Muse household. A more capable variation is currently in advancement.

Daily Debrief