Four months ago Meta announced it was done with open-source AI. This week, Mark Zuckerberg published a 6,500-word manifesto saying the opposite, and backed it with action.

On August 10, Meta released the open weights for Muse Glimmer, its 30-billion-parameter agentic model, and Zuckerberg confirmed the company will release open weights for Muse Spark 1.2 in the coming weeks. Muse Spark 1.2 is Meta’s most capable model and the one powering Muse Code, the terminal-based coding agent the company launched on August 5.

This is a genuine reversal. In April 2026, Meta’s introduction of proprietary Muse Spark marked what looked like a permanent break from the Llama open-source playbook. The company had assembled a new AI division, brought in Alexandr Wang from Scale AI, and chosen to keep weights closed. The message was clear: Meta was competing for frontier AI leadership and closed weights are how frontier labs protect their position.

The August 10 manifesto, titled “The Future is for Everyone,” walks that back.

Why Meta Changed Course

The competitive picture shifted faster than Meta expected. Alibaba’s Qwen team has released a string of high-quality open-weight models including Qwen 3.8, a 27-billion-parameter multimodal model that arrived on August 14. DeepSeek has continued building on its open release strategy. The result is that the open-weight model ecosystem has become substantially more capable without Meta’s participation.

Zuckerberg’s manifesto frames open-source AI as a geopolitical argument, not just a product one. He argues that US labs are operating under additional restrictions that Chinese labs don’t face, and that open-source models from US companies are necessary to maintain American leadership in AI. He describes open-source as “a positive and important force” for preventing AI capability from concentrating in too few hands.

A $1 billion “Future is for Everyone” fund will invest in communities where Meta operates data centers, but the more consequential announcement is the Muse Spark 1.2 open weights.

What Muse Spark 1.2 Actually Is

Muse Spark 1.2 launched on August 5, 2026, as a closed, API-only model priced at $1.25 per million input tokens. It was built for complex software engineering tasks, ships with a one-million-token context window, and integrates with persistent asynchronous background agents.

It is not a lightweight open-source model. It is the model Meta used to compete with Claude Code and OpenAI Codex. If the weights arrive as promised, enterprises will have access to a frontier-tier coding and reasoning model they can run on their own infrastructure, fine-tune on their own data, and deploy without per-token fees.

That is a meaningful shift in the enterprise AI landscape.

What This Means for Business

The most immediate practical implication is cost and control. Enterprises currently face a binary choice when they want capable AI in production: pay API fees to a cloud provider, or accept that open-weight alternatives are weaker. Muse Spark 1.2 as open weights would challenge that trade-off directly.

For organizations working with sensitive data, internal codebases, or regulated industries, the ability to run a top-tier model locally without data leaving the building changes the calculus on AI adoption. This is particularly relevant for finance, legal, and healthcare teams that have been waiting for open models to close the capability gap.

For data teams specifically, open weights mean fine-tuning. A Muse Spark 1.2 base model trained on your internal documentation, your data dictionary, or your product architecture will outperform a generic cloud model on your specific problems. That is the value case that Llama delivered for thousands of teams, and it looks like it is coming back at a higher capability level.

The broader story is about competition. When closed labs and open-weight models compete directly at the frontier, pricing pressure flows downstream. API costs drop, cloud compute costs drop, and the overall economics of running AI in production improve for businesses on both sides of the proprietary versus open-source debate.

Meta’s reversal is a sign that open-source AI has enough momentum that even the lab that briefly abandoned it felt the pull to return. The companies building on AI infrastructure should pay attention to that signal.

---

If you’re evaluating AI models for a specific business application and want to understand the trade-offs between managed API services and self-hosted open-weight models, the Omni Advisory team works through exactly these kinds of infrastructure decisions with clients.