Meta Superintelligence Labs Debuts Muse Spark, Its First AI Model After Zuckerberg's Billion-Dollar Overhaul

Meta debuted Muse Spark, a homegrown AI model it says significantly narrows the performance gap with models from OpenAI, Anthropic and others. The model, internally code-named Avocado and built over nine months by a team led by Chief AI Officer Alexandr Wang, is a major upgrade over its Llama 4 models. Meta shares rose 7% on the news.

Muse Spark is a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration, and is available at meta.ai and the Meta AI app. The model accepts voice, text, and image inputs but produces text-only output. It uses a fast mode for casual queries and several reasoning modes; a "shopping mode" highlights how Meta is embedding commerce-oriented AI directly into its consumer products. In the coming weeks, Meta says it will extend the model across WhatsApp, Instagram, Facebook, and Messenger.

The question looming over the launch is how much of this is genuine progress and how much is marketing positioning. A Meta executive described Muse Spark as competitive with the latest models from leading labs at specific tasks, including multimodal understanding and processing health information. In other areas, notably coding, the company openly acknowledges a gap between Muse Spark and the models already available. That is an unusual admission for a flagship product launch, and one that suggests Meta knows its new model still has ground to make up against GPT-5, Claude Opus 4, and Gemini 3.1 Pro.

The cost side of this bet is staggering. Meta entered a $100 billion multi-year agreement with AMD to deploy up to six gigawatts of AI infrastructure, deploying custom AMD Instinct MI450-based GPUs and 6th Gen AMD EPYC CPUs, with first gigawatt deployment shipments expected in the second half of 2026. Zuckerberg has said Meta plans to build tens of gigawatts this decade, and hundreds of gigawatts or more over time, calling the infrastructure buildout a strategic advantage. The $14.3 billion Meta invested to bring Wang aboard, through its acquisition of Scale AI, now rides directly on whether Muse Spark can deliver.

The organizational structure behind the model also marks a departure from how Meta has historically run AI projects. Meta Superintelligence Labs is the new team of expensive AI researchers helmed by Wang, and Zuckerberg has described its internal setup as "very flat," with none of the approval chains that defined prior AI work at the company. Meta Superintelligence Labs comprises four groups: TBD Lab managing large language models under Wang, FAIR for fundamental research, Products and Applied Research led by Nat Friedman for consumer integration, and MSL Infra led by Aparna Ramani for sustaining AI infrastructure. That structure places accountability squarely with Wang, who now carries the weight of Meta's most expensive corporate reorganization in years.

On the question of openness, Muse Spark represents a genuine strategic pivot. The model is currently closed, meaning its design and code will not be made public, marking a shift from Meta's prior open-source strategy. Before openly releasing versions of the new models, Meta wants to keep some pieces proprietary and ensure they do not add new levels of safety risk. Meta plans to release a version of Muse Spark under an open-source license eventually, but the timing remains vague, and the safety review process that must precede any release adds meaningful uncertainty to that timeline.

The release is being watched as a major test for Wang, who was brought in to fix the momentum issues Meta faced after its Llama 4 family failed to keep pace with the competition. The coding gap alone signals that Muse Spark is not yet a leap into undisputed frontier territory. But for a company that spent most of 2025 outpaced and reorganizing, launching a model that is competitive in multimodal reasoning, embedded across billions of daily users, and backed by a compute roadmap measured in gigawatts represents a credible opening statement in a race Meta had been losing.