Muse Spark Skipped Open Source, and Meta’s Safety Report Explains Why

Alexandr Wang called Muse Spark an “appetizer” at the Bloomberg Tech Summit in San Francisco on June 4, and the characterization was hard to dispute. The model scores 52 on the Artificial Analysis Intelligence Index, placing it fourth among frontier AI systems, behind Gemini 3.1 Pro, OpenAI’s GPT-5.4, and Claude Opus 4.6. Asked when the main course arrives, Wang, who leads Meta Superintelligence Labs (MSL), answered plainly: “We’re cooking it.”

Wang also confirmed at the summit that the model triggered internal safety alerts during development, with chemical and biological capabilities assessed as “high risk” in an unmitigated state, and that finding shaped the decision to keep the model’s weights proprietary. The Llama open-source ecosystem had accumulated 1.2 billion downloads by early 2026. The safety assessment gave Meta a concrete reason to close the door on public weight releases for its most capable system.

Wang’s Appetizer Defense

Muse Spark launched on April 8, the first major system from MSL since Zuckerberg assembled the group around Wang in June 2025, following Llama 4’s widely criticized April 2025 debut and the Behemoth model that Meta had been developing before the rebuild, delayed and ultimately never shipped. Nine months of work from scratch separated that low point from the April launch: new infrastructure, new architecture, new data pipelines.

At the summit, Wang’s case rested on pace. “We’re on a much faster trajectory because we’ve been doing all this work over the past year,” he said, naming data, compute, and research scaling as the three variables. The barrier to frontier-tier performance is “not money,” he told the audience, a claim set against a company projecting up to $145 billion in 2026 capital expenditures.

We’re cooking it. We’re seeing very exciting and promising results in the process of training it right now.

Wang, onstage at the Bloomberg Tech Summit in San Francisco on June 4, 2026, on when Meta’s next frontier model ships.

The model handles text, images, video, and audio natively, with a “Contemplating” reasoning mode positioned alongside Gemini Deep Think and OpenAI’s GPT Pro. It powers Meta AI across Facebook, Instagram, WhatsApp, Messenger, and Threads, reaching more than three billion combined users, and runs inside the Ray-Ban Meta smart glasses. Meta says the system achieves its reasoning performance at an order of magnitude less compute than Llama 4 Maverick, the previous mid-size flagship, through a training technique Wang’s team calls “thought compression”: during reinforcement learning, the model is penalized for extended reasoning chains, pushing it to solve complex tasks with fewer tokens. Wang flagged health capabilities, multimodal reasoning, and creative coding as the model’s strongest areas; consumer health AI at billion-user scale is where he expects capability gains to produce the most visible results.

Wang described Meta’s broader agent push as the long-term destination, building what he called “the best personal agents for everybody around the world.” A specific gap sits underneath that framing: reporting by The Information in May found that Meta’s internal shopping agent codenamed Hatch was still running on Anthropic’s Claude models, with a plan to switch to Meta’s own model at launch. Meta’s most prominent AI agent was running on a competitor’s model while Wang was describing the agent future onstage.

When the Model Triggered a High-Risk Alert

Before the April launch, Meta’s safety team ran the model through the company’s Advanced AI Scaling Framework. The assessment found that chemical and biological capabilities, in an unmitigated deployment, likely reached “high risk” under the framework’s thresholds. Mitigations brought residual risk to “moderate or lower,” and Meta published the 158-page Muse Spark Safety and Preparedness Report on April 28, three weeks after the model went live. Cybersecurity risk was assessed separately at “moderate or lower” without requiring the same mitigation depth.

Wang connected the biological finding directly to the open-source decision. When the company deploys a model inside its own products, he said, it has “a lot of ways to mitigate some of these risks.” Open-sourcing weights transfers both capability and risk to whoever downloads them, with no ability to enforce mitigations after that point. The control that a closed app deployment provides doesn’t exist for a freely downloadable weight file.

• 52 — Muse Spark’s Artificial Analysis Intelligence Index score, vs. 57 for GPT-5.4 and Gemini 3.1 Pro
• “High risk” — chemical and biological classification in unmitigated deployment, per the April 28 safety report
• 53 — Claude Opus 4.6’s score on the same index
• 18 — Llama 4 Maverick’s score on the same index before Wang’s nine-month rebuild

Wang said Meta hasn’t abandoned open-source AI. The company continues releasing models it judges safe for public weight distribution, with a safety review as the gate. Whether future Muse-series models clear that bar depends on how capable they are. The Llama brand is unresolved: “We have exciting debates about branding internally,” Wang said at the summit, “and nothing to share right now.”

The Llama Track and the Muse Track

April 8 produced an unusual simultaneous release. Meta shipped Llama 5 as an open-weight, community-licensed download and a closed proprietary system available only inside Meta’s apps on the same day. That dual release made the structural fork public in a way a sequential rollout wouldn’t have.

Llama 5 competes at the enterprise tier. On MMLU-Pro reasoning benchmarks, it scores 86.4 against GPT-5’s approximately 87. On LiveCodeBench coding tasks, it reaches 71.8 percent against GPT-5’s 70, a narrow lead. The Llama ecosystem had accumulated 1.2 billion total downloads by early 2026, averaging about one million per day, with US deployments at 35 percent of the total and Chinese models from Alibaba and DeepSeek absorbing 41 percent of downloads on Hugging Face by late 2025.

AI Index: Artificial Analysis Intelligence Index, April 2026. Llama 5 enterprise benchmarks per Meta and Epoch AI partial replication.

Llama’s free availability chips away at the commercial leverage of OpenAI and Anthropic, since a developer who can self-host a near-frontier model has less reason to pay API fees. Meta’s advertising revenue sits entirely outside that commodity dynamic. The closed track operates on different math: keeping architectural details and weights proprietary prevents competitors from extracting training signals from the model. Wang’s three-variable scaling argument (data, compute, research) functions only if the system being scaled stays in one place.

The geographic distribution of Llama downloads carries weight that Meta rarely publicizes directly. With 41 percent of Hugging Face Llama downloads coming from Chinese developers and model labs by late 2025, the open track has become part of an argument about AI infrastructure sovereignty that runs beyond individual developer preference. A company that provides the free base layer to global AI builders occupies a structurally different position than one that sells API access.

$145 Billion and a New Org Chart

The Spending Machine

Meta’s 2026 capital expenditure guidance runs from $125 billion to $145 billion, nearly double the $72.2 billion the company spent in 2025, targeting more than 1.3 million GPUs and roughly one gigawatt of AI computing capacity. Commitments include a $21 billion arrangement with cloud provider CoreWeave and a $27 billion data center joint venture in Louisiana with Nebius. Q1 2026 revenue came in at $56.3 billion with $26.8 billion in net income, giving Meta the cash flow to fund the build without immediate balance-sheet pressure.

The workforce cost of that reorientation landed on May 20. Meta notified approximately 8,000 employees of layoffs, representing 10 percent of its 78,865-person global workforce, while reassigning roughly 7,000 others to AI-focused roles organized into capability-based “pods.” Reality Labs, the Facebook social division, recruiting, sales, and global operations absorbed the heaviest cuts. MSL and AI infrastructure stayed in active hiring, with compensation packages for senior researchers reportedly reaching $100 million.

Internal Fault Lines

In March 2026, CTO Andrew Bosworth created a parallel Applied AI Engineering organization under Maher Saba, a Reality Labs veteran who reports to Bosworth rather than to Wang. Engineers across the company have been redistributed into both organizations, a dual structure insiders describe as a hedge against the risk that frontier model research doesn’t convert to commercial products quickly enough to justify the infrastructure spending.

The departure that preceded the whole reorganization was Yann LeCun, who led Meta’s Fundamental AI Research (FAIR) group for a decade before leaving in late 2025 after publicly calling Wang “young and inexperienced.” Within months, LeCun raised a $1.03 billion seed round for AMI Labs (Advanced Machine Intelligence Labs), backed by Nvidia, Bezos Expeditions, and Temasek, in what became Europe’s largest-ever seed round. The architectural argument LeCun took with him, that scaled large language models (LLMs) are the wrong path to general intelligence, remains active in the research community and his funding has given it a platform.

Wang founded Scale AI as a 19-year-old MIT dropout in 2016 and built it into a data-labeling company Meta valued at the implied price of $14.3 billion for a 49 percent non-voting stake. MSL’s four-group structure includes TBD Lab, Wang’s frontier model group; FAIR, now restructured; a products and applied research division under Nat Friedman, the former chief executive of GitHub; and an infrastructure unit under Aparna Ramani.

When Does the Entrée Arrive?

For developers, Wang’s summit appearance registered against a frustration that predates June 4. The model launched on April 8 without a public API. An application programming interface (API) is the only route for external developers to reach a proprietary model, the mechanism through which OpenAI and Anthropic generate developer revenue. Wang told developers an API would arrive “soon” at launch. As of June 2, The Wall Street Journal reported that Meta had no launch date set and had repeatedly pushed the timeline.

A company spokesman confirmed to Reuters that the interface was in partner testing and would ship during June. Developers outside that cohort have had no access to the model since April, no open weights to run locally and no API to query remotely. For the community that ran Llama at one million downloads per day precisely because it required no gatekeeper approval, the model that followed it has been functionally invisible for nearly two months.

Zuckerberg told investors on the Q1 2026 earnings call that companies ask Meta every week about setting up an API service, and that a cloud-computing business is “definitely on the table” as a way to monetize spare GPU capacity. That commercial track depends on the developer API being in production. Meta’s spokesman confirmed the interface was in partner testing as of June 5, with no public launch date set.