WORLD

Meta's superintelligence team launches first AI model, taking on OpenAI and Google

by
Kim Young-chul
Published : April 9, 2026 - 07:21:30
Updated : Nov. 20, 2026 - 13:22:13
    • Copy Completed!

View Korean Original

Unlike open-source Llama, new Muse series released as closed model; rollout planned for social media platforms and AI glasses

The Meta logo. [AFP]
The Meta logo. [AFP]

A research team Meta assembled through an aggressive talent recruitment drive last year has unveiled its first AI model, one that rivals offerings from the company's major competitors.

Meta said its Meta Superintelligence Labs (MSL), led by Chief AI Officer Alexander Wang, released "Muse Spark" — the first model in its Muse series — on Wednesday. The model had been developed under the codename "Avocado" before its public launch.

Meta described Muse Spark as "designed to be small and fast, but capable enough to reason through complex questions in science, math and health."

Benchmark scores released by Meta show Muse Spark matches or surpasses rivals including OpenAI's GPT-5.4, Google's Gemini 3.1 Pro and Anthropic's Claude Opus 4.6 on several measures.

Muse Spark posted the highest score among the compared models on CharXiv Reasoning, a chart comprehension benchmark, at 86.4 percent. It also scored 80.4 percent on MMMU Pro, a multimodal recognition benchmark, putting it on par with competing models.

On coding benchmarks, Muse Spark scored 77.4 percent on SWE-bench Verified and 52.4 percent on SWE-bench Pro — slightly below the three rival models, though the gap was not large.

Meta said Muse Spark will also feature a "Contemplating" mode, in which multiple agents reason simultaneously to tackle complex problems. The company said this would allow the model to compete with dedicated high-performance reasoning models such as Gemini 3.1 Deep Think and GPT-5.4 Pro.

With Contemplating mode enabled, Muse Spark scored 50.2 percent on the Humanity's Last Exam (HLE) benchmark, which measures expert-level reasoning — topping Gemini 3.1 Deep Think's 48.4 percent. However, it trailed Claude's recently previewed Mythos Preview, which scored 56.8 percent and has so far been made available only to select companies.

AI performance analytics firm Artificial Analysis gave Muse Spark an aggregate score of 52, placing it fourth behind Gemini 3.1 Pro Preview (57), GPT-5.4 (57) and Claude Opus 4.6 (53). That still marks a significant improvement over Llama 4 Maverick, Meta's most recent model before the superintelligence team was formed, which scored just 18.

Meta said Muse Spark also performed well on safety evaluations, refusing 98 percent of high-risk requests related to biological weapons — higher than Claude Opus (95.4 percent), GPT (74.7 percent) and Gemini (61.5 percent).

While Meta released its Llama series as open-source models, the new Muse series follows the same closed approach as its major rivals. Meta said it may consider making future versions open-source.

Muse Spark is available starting Wednesday through the meta.ai website and app.

Meta plans to integrate Muse Spark into its social media platforms — Facebook, Instagram, Threads and WhatsApp — as well as its AI glasses.

After the Llama series drew underwhelming reviews last year, Meta poured massive resources into building MSL, including recruiting Scale AI CEO Alexander Wang in a deal valued at $14.3 billion (about 21 trillion won), and has since focused on developing a new generation of AI models.


yckim6452@heraldcorp.com
This content was produced with the assistance of AI translation services.

MOST READ