Meta Open Source AI: Muse Spark and the Fight for Global Access

What Muse Spark Actually Does

Muse Spark is not a minor upgrade. Meta rebuilt its entire AI system over nine months. The result? A model that matches its older Llama 4 in skill — but uses just a tenth of the computing power (compute). That matters hugely at Meta's scale. The company runs AI on WhatsApp, Facebook, Instagram, and Messenger — reaching over three billion users every day.

The model introduces three distinct modes of operation. Instant mode handles quick answers to simple questions. Thinking mode tackles multi-step reasoning tasks such as analysing legal documents or extracting nutritional information from product photos. Contemplating mode runs multiple AI agents at the same time. It is built to rival Google's Gemini Deep Think and OpenAI's GPT Pro (Generative Pre-trained Transformer).

Where Muse Spark leads the industry is health. On HealthBench Hard — a test for health question answers — Muse Spark scores 42.8. That puts it well ahead of Gemini 3.1 Pro (20.6), GPT-5.4 (40.1), and Grok 4.2 (20.3). Meta worked with over 1,000 doctors to build these features. The goal? Making AI tools genuinely useful for everyday people — not just developers and big firms.

The Open Source Gamble: Why Meta Changed Course

For years, Meta was the undisputed champion of Meta open source AI. Its Llama models reached 1.2 billion downloads by early 2026, averaging roughly one million downloads per day. Developers, researchers, and startups worldwide built products on Llama. No other company could match that reach.

Then Llama 4 disappointed. The model family, released in early 2025, failed to captivate developers and fell significantly behind competitors. CEO Mark Zuckerberg responded fast. He brought in Alexandr Wang, spent $14.3 billion, and set up Meta's new AI lab with one goal: rebuild everything from scratch.

The result is a hybrid strategy. As Axios reported, Meta plans to release open source versions of its new models later. But some of its biggest models will stay closed. Wang calls this move pragmatic, not a retreat. Meta wants to stay open enough to keep developer trust. But it will guard its edge where it counts.

This approach mirrors a broader industry shift. Alibaba, once a champion of open source AI through its Qwen models, recently launched closed-source models focused on profit. China's top AI labs are going closed at the same time Meta pulls back. The global open source movement is losing its biggest backers — right when it matters most.

What This Means for Developers in the Global South

The technology democratization angle is the part of this story that most tech coverage misses entirely. When Meta releases an open source AI model, the primary beneficiaries are not Silicon Valley startups with venture capital funding. They are developers in Lagos, São Paulo, Jakarta, and Nairobi. These builders cannot afford $20 per million tokens (units of text AI processes) from OpenAI's API (application programming interface). But they can download and run a Llama model on their own hardware.

This is not theoretical. A recent International Labour Organization (ILO) study looked at 135 countries. It found that workers in developing nations have enough internet access to be replaced by AI. But they lack the digital systems to benefit from it. Open source models are one of the few ways to close this gap. They let people run AI locally — with no need to pay for cloud services run by American and Chinese giants.

The World Economic Forum has documented how countries across the Global South are building language-first AI models and investing in public compute infrastructure. Open source foundations like Llama make this possible. Without them, the cost of entry becomes prohibitive for nations that most need AI-powered economic growth.

Demand for cheaper AI options is already reshaping markets. Networks offering pro-grade AI at fair prices are winning over professionals who need strong tools but lack big budgets.

If Meta restricts its best models to proprietary channels, billions of people lose access to the most powerful free AI tools on the planet. The 40 occupations already at high risk of AI replacement largely employ people in developing economies. Keeping frontier models open is not just a technical preference — it is an economic justice issue.

"Nine months ago, we rebuilt our AI stack from scratch. New infrastructure, new architecture, new data pipelines. This is step one. Bigger models are already in development with plans to open-source future versions."
— Alexandr Wang, Chief AI Officer, Meta

The $325 Billion Question: Who Controls AI's Future?

The global generative AI market could grow from about $22 billion in 2025 to nearly $325 billion by 2033, says Grand View Research. That is a growth rate of over 40 percent each year. At that scale, one question decides everything: are the best AI models open or closed? The answer shapes who gets to join the biggest economic shift since the internet.

Right now, three players dominate. OpenAI and Anthropic sell to big firms and governments. Meta's open source path has served the long tail — small developers and lean startups. As Geoffrey Hinton warned the UN, the critical question is not whether AI advances but who gets to steer it. If only a few firms control the best AI, it becomes a tool for keeping power — not sharing it.

Meta's Muse Spark launch reveals the tension at the heart of this debate. The company's stock rose approximately 6.5 percent on launch day, suggesting Wall Street rewards proprietary control over open distribution. Yet Meta's edge — three billion users across its apps — rests on one thing: giving AI to everyone, not just those who pay.

The upcoming months will be decisive. OpenAI and Anthropic are both preparing new flagship models. If Meta releases open source versions of Muse quickly enough, it reinforces its position as the engine of accessible AI. If it stalls — or weakens the open versions — developers will look elsewhere. And the open AI movement will lose its strongest ally.

How to Think About Meta Open Source AI Going Forward

For developers, small business owners, and anyone building with AI tools, here is what the Muse Spark launch means in practical terms:

The Llama ecosystem is not dead. Meta has not discontinued Llama. The 1.2 billion downloads represent an active, thriving community. Open source versions of Muse-series models are promised, though no firm timeline has been given.

Proprietary does not mean inaccessible. Meta is embedding Muse Spark directly into WhatsApp, Facebook, Instagram, and Messenger — free services with global reach. A farmer in rural India using WhatsApp will access the same AI as a marketing executive in London. The model is proprietary in terms of developer access, but universally available through Meta's consumer products.

The hybrid approach may actually work. Meta keeps its strongest models closed. It releases smaller open source versions alongside them. This way, it can fund its huge R&D costs ($115–135 billion in 2026 alone) and still feed the open source world. This is the model that makes AI accessible to the widest audience. Not through charity — but through a business that funds both new ideas and openness.

Watch the timeline. The key measure is not what Meta promises. It is how fast the open versions arrive after the closed launch. A three-month gap is sustainable. A twelve-month gap would effectively kill developer trust.