Meta has officially abandoned the open-source AI philosophy that defined its brand, launching Muse Spark as a closed-source proprietary model for the first time in company history. The shift represents a dramatic reversal for the organization that created Llama, the open-weight model that sparked the entire open-source AI movement. Developed by Meta's newly assembled Superintelligence Labs under leadership from Alexandr Wang, Muse Spark launches with no public weights, no open license, and no commitment to releasing the source code .

Why Did Meta Abandon Open-Source AI?

The answer lies in competitive pressure and past failures. Meta's previous Llama models largely failed to gain traction, and the company faced a major scandal last year when it allegedly faked benchmark results for Llama 4 to make the model appear more capable than it actually was. Former Meta AI head Yann LeCun, who left the company amid the controversy, told the Financial Times that the results "were fudged a little bit," and that Mark Zuckerberg "basically lost confidence in everyone who was involved in this" .

To recover, Meta spent hundreds of millions of dollars assembling a new team from scratch over nine months. The result is Muse Spark, which performs exceptionally well on health-related reasoning tasks. On HealthBench Hard, a specialized benchmark for medical knowledge, Muse Spark scored 42.8, beating OpenAI's GPT-5.4 (which scored 40.1) and Google's Gemini 3.1 Pro (which scored 20.6). The model ranks fourth overall on the Artificial Analysis Intelligence Index, behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 .

Yet despite these strong results, Meta is keeping the model proprietary. The company plans an open-source version eventually, but without a specific date, which in practice means the competitive advantage stays locked away for now .

What Can Muse Spark Actually Do?

Muse Spark accepts multiple types of input and operates in three distinct modes designed for different tasks. The model is built for speed and efficiency, with a small footprint that allows it to run across Meta's entire ecosystem of products .

• Input Types: The model accepts voice, text, and image inputs, with text-only output for now, allowing users to interact with the AI in multiple ways.
• Fast Mode: Designed for quick queries that don't require deep reasoning, enabling rapid responses to straightforward questions.
• Contemplating Mode: For complex tasks, the model runs sub-agents in parallel to break down problems and provide more thorough analysis.
• Shopping Mode: A specialized version that layers behavioral signals on top of the base model to assist with purchase decisions and product recommendations.

The model is already live in the Meta AI app and meta.ai, with rollout planned for WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban smart glasses in the coming weeks, starting in the United States .

How Does This Compare to Competitors?

Meta's decision to go closed-source puts it in direct competition with OpenAI, Anthropic, and Google, which have all pursued proprietary models. However, the company still faces a credibility gap. An executive told Bloomberg that Muse Spark "likely won't be able to keep up with competing models," and Meta itself acknowledged in a blog post that the model "is an early data point on our trajectory, and we have larger models in development" .

Meanwhile, Google released Gemma 4, an open-source model with 31 billion parameters that beats Meta's Llama 4 across every major benchmark. Gemma 4 scored 89.2% on the AIME 2026 Math benchmark compared to Llama 4's 88.3%, and achieved 80% on LiveCodeBench v6 versus Llama 4's 77.1%. Google released Gemma 4 under the Apache 2.0 license with no restrictions, no fees, and no gating, making it freely available to anyone .

"Meta open-sourcing Llama was a strategic move to commoditize competitors and commoditize compute. Muse Spark ships closed-source, the first time Meta has ever done that. An open-source version is planned but without a date, which in practice means the competitive advantage stays proprietary for now," noted an industry analyst.

Industry Analysis, The Creators' AI Edition

What Does This Mean for Meta's AI Strategy?

The closed-source shift signals that Meta is finally taking AI seriously as a competitive battleground. The company has struggled to stay relevant in the rapidly evolving AI landscape, making headlines for legal liability issues rather than technological breakthroughs. By keeping Muse Spark proprietary, Meta is betting that competitive advantage matters more than the goodwill that comes from open-sourcing technology .

However, questions remain about whether Muse Spark can actually jumpstart Zuckerberg's goal of playing in the big leagues. OpenAI, Anthropic, and Google have already built powerful coding assistants and secured lucrative enterprise customers. Meta's model, while strong on health reasoning, still needs to prove itself across a broader range of tasks .

For now, Muse Spark is free for all users. However, Meta executives told Bloomberg that the company is considering paywalling the model behind a subscription in the future, following the playbook of competitors like OpenAI .