Meta's AI shift: from open Llama to closed Muse, and what it costs

Meta has moved its newest AI models, Muse Spark in April 2026 and Muse Wonder in October 2026, away from the open release approach that defined its Llama line. Muse Spark was closed, with Meta saying it hopes to open future versions. Meta guides to $130 billion to $145 billion of capital spending in 2026 to support the work.

This post separates what Meta has stated in its own releases from what outlets report, explains what "open" and "closed" mean for a customer, and lists what is still unknown. Nothing here is investment advice or a prediction.

Key numbers at a glance

Figures are in US dollars. Company figures come from Meta's second quarter 2026 results release (quarter ended 30 June 2026, filed 29 July 2026). Where a figure comes from a news report, it says so.

• Revenue, Q2 2026: $60.80 billion, up 28 percent year over year. Source: Meta results release.
• Total costs and expenses, Q2 2026: $42.03 billion, up 55 percent. That includes $2.40 billion of legal charges and $1.18 billion of severance. Source: same release.
• Full year 2026 total expense outlook: $165 billion to $169 billion. Source: same release.
• Capital expenditure, Q2 2026: $31.08 billion including principal payments on finance leases. Source: same release.
• 2026 capex guidance: $130 billion to $145 billion, narrowed from $125 billion to $145 billion. Source: same release.
• 2025 capex: $72.22 billion, as reported by TheNextWeb on 8 April 2026.
• January 2026 capex guidance: $115 billion to $135 billion, as reported by the same outlet.
• Headcount at 30 June 2026: 75,472, down 1 percent year over year. About 8,000 employees were affected by a May 2026 reduction. Source: Meta results release.
• Scale AI investment: $14.3 billion for a 49 percent non-voting stake, as reported by TheNextWeb.
• Muse Spark launch: 8 April 2026, closed source, per TheNextWeb.
• Muse Wonder and Muse Wonder Full: announced 12 October 2026 (updated 13 October) by Meta.
• 2027 capex guidance: not given in the release.

From Llama to Muse: the timeline

For most of 2023 to 2025, Meta released powerful Llama model weights for others to download, an unusual choice for a company of its size. TheNextWeb's April 2026 report says the Llama series "set the template" for open source AI model development through 2025.

In June 2025 Meta announced a new group, Meta Superintelligence Labs, usually shortened to MSL. TheNextWeb reports that Alexandr Wang, who had run the data company Scale AI, was named Meta's first chief AI officer, and that Meta invested $14.3 billion for a 49 percent non-voting stake in Scale AI. Wang wrote on X, according to that report, that Meta had "rebuilt our AI stack from scratch" over nine months.

On Wednesday 8 April 2026, the first MSL model, Muse Spark, arrived. The report says it was known internally as Avocado and had been delayed earlier in the year after trailing rivals in internal tests. It is a closed model, a break from Llama. Meta said it hopes to release future versions under an open source license. As of 19 October 2026, Meta has not published a later statement that confirms or dates any such release.

Then on 12 October 2026 Meta announced Muse Wonder and Muse Wonder Full, which it calls "our most capable models to date." They build on Muse Spark. Muse Wonder is available to developers through the Meta Model API and through a coding product called Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Muse Wonder will soon power experiences in the Muse assistant, rolling out first to subscribers.

The announcement does not say whether the weights are open or closed. Since the models are offered through an API and a subscription product, and no download is mentioned, the plain reading is that they are served by Meta rather than downloaded. That is an inference; Meta's post does not say so.

What "open" and "closed" mean for a customer

An open weights model is one where the trained numbers are published, so you can download and run the model yourself. A closed model is only available through the owner's service. For the formal differences between open weights and open source, see the explainer on that definition.

Here is what changes for a buyer.

• Where your data goes. With an open model you can run it in your own cloud account or on your own hardware, so prompts never leave your control. With a closed model you send data to the provider and rely on its contract and controls.
• Who sets the price. An open model has a cost you can estimate from hardware. A closed model has a price list that the provider can change.
• How long it stays available. Downloaded weights keep working. A hosted model can be retired or changed.
• Who fixes problems. A downloaded model is your responsibility to patch and monitor. A hosted model is the provider's.
• What you can inspect. With open weights you can test, fine tune and examine the model. With a closed model you can only test its behavior.

None of this makes one choice right. A team with strict data residency rules may prefer open weights. A small team with no GPU skills may prefer a hosted service. The point is that Meta's shift moves one large source of open models toward the hosted side, which matters if you built a plan around Llama.

If you already use Llama, the models you downloaded still run. Meta has not announced that it has withdrawn them. The question is what comes next, and for that Meta's public statements are thin: it hopes to open future versions, and it has not committed to a date.

How good are the models, and who says so

TheNextWeb's April report lists Meta's own published results for Muse Spark. On the Artificial Analysis Intelligence Index version 4.0, Muse Spark scored 52, which put it fourth behind Gemini 3.1 Pro Preview and GPT-5.4 at 57 each and Claude Opus 4.6 at 53. On GPQA Diamond, a graduate level science test, it scored 89.5 percent against 94.3 for Gemini 3.1 Pro. On SWE-bench Verified, a software engineering test, it scored 77.4 percent. It led on some tasks, for example CharXiv Reasoning at 86.4 and HealthBench Hard at 42.8 percent. These are scores Meta published, relayed by a news outlet, not independent tests.

For Muse Wonder, Meta's post gives no scores. It claims "state-of-the-art performance on personal and professional agentic tasks," strength in law, finance and health, and "frontier coding performance." It describes two features: Active Memory, which carries relevant information across conversations, and Active Focus, which keeps working toward a goal, tries other approaches after errors and asks for clarification. Meta says it trained the models to avoid sharing private information without permission and promised a full safety evaluation report "soon." As of 19 October 2026 Meta has not published it.

The practical advice is the same as for any model: wait for independent evaluations and run your own tests on your own tasks before relying on a vendor's claims.

Where the money goes

Meta's second quarter results make the scale clear. Revenue was $60.80 billion, up 28 percent. Capital expenditure was $31.08 billion in the quarter, which is more than half of the revenue. Total costs and expenses rose 55 percent to $42.03 billion, partly because of two one time items: $2.40 billion of legal proceedings charges and $1.18 billion of severance from the May 2026 workforce reduction.

Capital expenditure, or capex, is spending on long lived assets such as data centers, servers and networking. Meta's 2026 capex guidance is now $130 billion to $145 billion. The company narrowed it from $125 billion to $145 billion given in April. TheNextWeb reports an earlier January range of $115 billion to $135 billion and a 2025 total of $72.22 billion. Set against the reported 2025 figure, the guidance range puts the 2026 plan roughly 80 to 100 percent higher than 2025. That is a simple calculation from a reported figure and a company range, so treat it as approximate.

The release does not give 2027 capex guidance. It does give a full year 2026 total expense outlook of $165 billion to $169 billion. Headcount at the end of June was 75,472, down 1 percent from a year earlier, and Meta said about 8,000 employees were affected by the May reduction, most of whom would leave the count by the end of the third quarter. So the company is spending much more on machines while shrinking its payroll.

Meta's chief executive, Mark Zuckerberg, said in the release that "AI is accelerating our core business today" and that he is "optimistic about the potential ahead." The release gives no further detail on how much of the spending is for MSL research compared with serving advertising and recommendation systems for its apps. Meta does not disclose that split.

Spending that does not appear as capex

Meta also builds some capacity through partners. The capex scoreboard post describes a Louisiana campus held by a joint venture, where Meta is the tenant. That description relies on secondary reports, so check the original filings if the detail matters to you. The general point stands: a company can use capacity without recording the building as its own capex, so reported capex is a floor, not a ceiling.

Meta's own newsroom gives only a little on infrastructure. A post dated 7 October 2026 is an interview with Santosh Janardhan, Meta's head of infrastructure, on how Meta powers its AI, what a gigawatt is, how it chooses chips and why it builds its own data centers. The interview gives no figures. The newsroom also lists related headlines about partnerships with AWS on Graviton chips and with Arm on data center silicon; their terms are outside the scope of this post.

A gigawatt is a billion watts of electrical power, roughly the output of a large power station. Data center capacity is now counted in gigawatts because power, not floor space, is often the limit. For more on cost per gigawatt, see the explainer on data center cost per megawatt and gigawatt.

How to read a lab's announcement

Announcements from AI labs follow a pattern, and knowing it helps you read Meta's posts with the right level of trust. First comes a claim of capability, such as "most capable models to date." Second comes a list of features with friendly names. Third comes availability, which tells you who can actually use the thing today. Fourth, sometimes, comes evidence: scores, test names and a technical report.

The Muse Wonder post has the first three and promises the fourth later. That is not unusual, and it is not a criticism. It does mean that on 19 October 2026 the capability claims are statements, not findings. A careful customer waits for the safety report and for outside testers, and meanwhile runs a small trial on real tasks. A trial of a week or two, with a fixed set of tasks and a written score sheet, tells you more than any launch post.

It also helps to check the availability line against your needs. A model offered only through a developer API and a subscription product is a service. You will need a contract, a data handling review and a plan for what happens if the provider changes the model or its price. A model with downloadable weights is closer to software you install, and then your own security checks apply to the files and the serving tools. Meta's post describes the first kind of access for Muse Wonder.

What is firm and what is uncertain

Firm, from Meta's own releases:

• The Q2 2026 revenue, expense, capex and headcount figures, and the 2026 capex and expense guidance.
• The 12 October 2026 announcement of Muse Wonder and Muse Wonder Full and the channels through which they are offered.
• Meta's statement that a full safety evaluation report is coming.

Reported by an outlet, not confirmed in Meta's own releases:

• That Muse Spark is closed and was internally called Avocado.
• The Scale AI stake of $14.3 billion for 49 percent.
• The benchmark scores for Muse Spark.
• The 2025 capex of $72.22 billion and the January 2026 guidance range.

Unknown:

• Whether Muse Wonder weights will ever be published, and under what license.
• The split of capex between research, ad systems and other uses.
• Independent test results for Muse Wonder.
• Meta's 2027 spending plans.

What to watch next

• Meta's third quarter results. They are not out as of 19 October 2026; check Meta's investor site for the date. Watch the 2026 capex range and any first view of 2027.
• The Muse Wonder safety report. Meta says it is coming "soon."
• Any open release. Meta said in April it hopes to open future versions. Watch for a named model, a date and a license.
• Independent evaluations. Compare them with the claims.
• Pricing for the Meta Model API. The Muse Wonder post does not include a price list.
• Chip partnerships. The Arm and AWS headlines suggest Meta is widening its chip sources. The details are worth reading when you plan around supply.

Quick check

If you want a record of which components your own software depends on, Safeguard offers SBOM and dependency scanning, and a free scan is available at https://gold.safeguard.sh/scan.

Sources

• Meta, "Meta Reports Second Quarter 2026 Results": https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx
• Meta, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents," 12 October 2026: https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
• Meta, "Why Data Centers Are Such a Big Part of Meta's AI Approach," October 2026: https://about.fb.com/news/2026/10/meta-data-centers-ai-approach/
• Meta newsroom listing, checked 19 October 2026: https://about.fb.com/news/
• TheNextWeb, report on the Muse Spark launch, 8 April 2026: https://thenextweb.com/news/meta-muse-spark-msl-first-model
• SEC EDGAR list of Meta 8-K filings, checked 19 October 2026: https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001326801&type=8-K&dateb=&owner=include&count=10

Last reviewed: 19 October 2026.