Meta’s Llama Just Went Fully Autonomous. Here’s What Happened in the First 72 Hours

This narrative can be told in a way that makes perfect sense. Reviews of Llama 4, which debuted in April 2025, ranged from passive to outright hostile. Internal dissatisfaction was coming out through the typical avenues, such as blogs from disgruntled developers, anonymous sources, and the kind of subtle institutional deflation that large corporations seldom publicize but are unable to completely conceal.

According to some accounts, Mark Zuckerberg was dissatisfied. Thus, he paid $14.3 billion for a 49 percent stake in Scale AI, appointed Alexandr Wang, the company’s co-founder and CEO, as Meta’s first-ever chief AI officer, and instructed a new team to rebuild the entire stack from scratch—all with the particular decisiveness that has defined his larger strategic moves. The results were announced on April 8, 2026, nine months later. Spark the Muse. exclusive. locked. Not Llama.

Important Information

On paper, the model is quite excellent, and the benchmark results are optimistic even though they haven’t been independently verified at full scale yet. Muse Spark is ranked fourth in the world on the Artificial Analysis Intelligence Index, only after Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Placing fourth in the world is a reasonable outcome for a business that, as recently as six months ago, was considered an also-ran in AI discussions.

Although it’s important to remember that Meta has already been discovered altering benchmark statistics, the early third-party testing is holding up better than some observers anticipated. The concept incorporates what the business refers to as multi-agent orchestration, which allows it to plan and carry out multi-step operations instead of only reacting to individual prompts. It is natively multimodal, encompassing text, graphics, structured data, and tool use. The most obvious difference between it and the Llama family it is replacing is that.

During the first 72 hours, several different reactions occurred concurrently. First to move were investors. In response to what the market perceived as a reliable indication that the company’s $115–$135 billion AI capital investment plan for 2026 may actually yield benefits worth the price, Meta’s shares increased by almost nine percent in the days after the disclosure. Analysts who were concerned about returns had been quietly alarmed by that amount of spending, which was almost twice as much as the previous year.

Although Muse Spark did not completely allay that concern, it did change the topic of discussion from “can Meta compete” to “where does Meta fit in a four-way race,” which is a new and more palatable question. There was a sense that investors had been waiting for approval to believe in the rebuild as they saw the stock move in real time. They got that from the launch.

The developer community was far more intricate. With 1.2 billion downloads and over a million new downloads per day, Llama had developed into something truly unique in the AI sector: a free, open-weight model on which actual companies were creating real goods. The openness of Llama—the capacity to self-host, optimize, and deploy without paying per-token to a proprietary API—was particularly appealing to a sizable percentage of those users.

According to one analysis, self-hosting Llama was 88% less expensive than commercial options. That is not a small difference. For many small teams and freelance developers, this is what makes a project feasible or unviable. That’s not what Muse Spark provides. Although Wang confirmed the change on social media and stated that future iterations of the model would eventually be open-sourced, the developer community has heard similar claims in the past, so skepticism is totally justified in this case.

Although they might not see the logic behind the changes right away, most users outside of the development community are likely to discover the autonomous capabilities first. With a deployment to Facebook, Instagram, WhatsApp, and Meta’s AI glasses in progress, Muse Spark is already operational on the Meta AI app and on meta.ai. This is significant since billions of people are reached every day by those platforms combined.

With over two billion users on WhatsApp Business alone, industry commentators are characterizing Muse Spark’s implementation there as Meta’s attempt to create an autonomous digital layer integrated into commercial messaging, similar to what WeChat has long done in China. In addition to Muse Spark, Meta is incorporating technology from Manus AI, an autonomous agent firm based in Singapore that it purchased in late December 2025. Manus AI manages the business-facing aspect of the agent ecosystem.

The first 72 hours also resulted in a noticeable increase in the amount of AI-generated content that flooded social media platforms; this is the type of low-quality, algorithmically-pumped output that has begun to accumulate its own shorthand. This was less covered in the official coverage. This was mentioned as an early indication of what completely autonomous, user-deployable agents of Meta’s size would signify for the information environment in the upcoming months in some quarters of the AI research community.

To say how significant that worry is is still premature. Strong rejection behavior in high-risk domains and a declared lack of autonomous capabilities in cybersecurity danger scenarios are highlighted in Meta’s own safety documentation for Muse Spark. The first 72 hours were unable to provide an answer to the question of whether those safeguards hold at scale, over billions of deployment instances. That test has a longer duration. The outcomes are still being received.