1.2 billion downloads. That's how many times Meta's open Llama models had been downloaded by the end of April 2025.
Then Muse showed up.
Since April 2026, Meta's strongest language model has been called Muse Spark, and Meta no longer publishes its weights. Under the Muse name you'll now also find an image model, an announced video model, an open model, a coding agent, and, since September, a personal AI agent that's simply called "Muse." Keeping track of all that takes some doing.
This article is my attempt to sort it out: user numbers, revenue, investment, Meta Superintelligence Labs, Llama, and the full Muse family with benchmarks, pricing and controversies, all in one place. For how Meta stacks up against the rest of the industry, check my AI statistics.
- Meta AI hit 1B monthly users in May 2025, counted across Facebook, Instagram, WhatsApp and Messenger combined
- For 2026, Meta expects capital expenditure of $130-145B, up from $72.22B in 2025
- Six different things now run under the Muse name, from the Muse Spark language model to the open Muse Glimmer 30B to the personal Muse agent
- Muse Image dropped from rank 2 at launch to rank 7 on Arena.ai, and Muse Video still isn't publicly available
1. Meta AI at a glance
Meta isn't a pure AI lab like OpenAI or Anthropic. The company makes its money from advertising and pours a growing share of it into AI, from the Meta AI assistant to its own data centers to the Muse models. The key facts:
As of September 2026. Sources: Meta Investor Relations, Meta Newsroom, CNBC.
In the second quarter of 2026, $59.36B of Meta's $60.80B in revenue came from advertising. That business is exactly what funds Meta's multi-billion-dollar AI bet.
2. Sorting out the Muse family
Before we get to the numbers, it's worth looking at the names. Meta uses "Muse" for models, tools, and a consumer product all at once. The table below untangles the six different Muse things:
As of September 2026. Sources: Meta Newsroom, research.meta.ai, TechCrunch.
The most common mix-up is between the Muse agent and the Muse Spark model. The agent is its own consumer app; Muse Spark is the language model it runs on.
And then there's Meta AI, the assistant inside Facebook, Instagram, WhatsApp and Messenger. Meta says it rebuilt the assistant and connected it to Muse Spark.
3. How many users does Meta AI have?
In September 2024, Mark Zuckerberg told Meta Connect that Meta AI had nearly 500 million monthly active users. At the time, the assistant hadn't even launched in some major markets, including the EU.
At the May 2025 shareholder meeting came the 1 billion mark. That sounds like a record, but it needs some context.
Since then, Meta hasn't given a new total. Instead, Zuckerberg talks about engagement. Since the Muse Spark rebuild, daily users of the assistant are up 60% according to the Q2 2026 call. There's no clean time series, just these data points:
Figures stated by Mark Zuckerberg (Meta Connect 2024, 2025 shareholder meeting, Q2 2026 call).
A standalone Meta AI app has existed since April 29, 2025. In Europe, Meta AI launched on March 19, 2025 in 41 countries and 21 overseas territories. At launch, the region only got a text model that wasn't trained on data from EU users.
For how AI usage overall is developing, see my article How many people use AI?
4. Revenue, advertising and AI in the ad business
Meta's AI business runs mostly through advertising. Better models target ads more precisely, and AI tools create the ad creative itself. The numbers from the second quarter of 2026:
Q2 2026, published July 29, 2026. Sources: Meta Investor Relations, earnings call.
Meta sees its biggest lever in Advantage+, its automated campaign management. According to CFO Susan Li, the end-to-end solutions hit an annual revenue run rate of over $75B in the second quarter, meaning current revenue extrapolated to a full year.
Advertisers are buying in too. By the end of 2024, more than 4 million advertisers were using at least one of Meta's generative AI ad tools, up from 1 million just six months earlier. By the first quarter of 2026, Meta said that number had passed 8 million.
The effect shows up in the ads themselves. Better models brought 8.3% more ad clicks and 15.7% more conversions on Facebook, according to Susan Li. On Instagram, early pilots with language models lifted app conversions by 1%.
For the full year 2025, Meta's revenue hit $200.97B, up 22% year over year, with net income of $60.46B.
5. Capex: how much Meta invests in AI
Capex stands for capital expenditure: money spent on data centers, servers and networking gear. Under current guidance, this line item at Meta will more than triple between 2024 and 2026:
In January 2025, Zuckerberg guided for $60-65B in capex for the year. It ended up at $72.22B.
For 2026, Meta has already revised that guidance twice:
Capex including principal payments on finance leases. Source: Meta Investor Relations.
That last revision wasn't quite as simple as it looks, though. Meta itself calls it a narrowed range: only the lower end moved up to $130B, while the upper end stayed the same.
The quarter-over-quarter comparison shows just how fast the pace is picking up. In the second quarter of 2026 alone, Meta spent $31.08B, compared to $19.84B in the first quarter.
What surprises me most is the contrast with headcount. At the end of 2025, Meta employed 78,865 people. By the end of June 2026, that was down to 75,472. The capital is flowing into data centers, not new jobs. For comparison, check my OpenAI statistics.
6. Meta Superintelligence Labs (MSL)
In summer 2025, Zuckerberg reorganized Meta's AI work. On June 30, 2025, he announced Meta Superintelligence Labs, MSL for short, in a memo. It brings foundational research, products and infrastructure under one roof:
Sources: CNBC, SiliconANGLE, Meta.
The deal with Scale AI laid the groundwork. Meta paid $14.3B for a 49% non-voting stake, and Scale founder Alexandr Wang moved to Meta as Chief AI Officer.
Alongside Wang, former GitHub CEO Nat Friedman and Daniel Gross round out the leadership. Gross, previously CEO of Safe Superintelligence, handles technical architecture and capacity planning at MSL.
The reorganization wasn't entirely smooth. In October 2025, Meta cut around 600 roles in its AI division, leaving MSL with just under 3,000 employees. In November 2025, Yann LeCun left Meta after twelve years leading FAIR to start his own venture focused on so-called world models.
MSL's first model arrived in April 2026. It's called Muse Spark.
7. Llama: 1.2 billion downloads
Before Muse, Llama was Meta's flagship. The open model family added hundreds of millions of downloads in just a few months:
The 1 billion mark fell on March 18, 2025. By LlamaCon in late April, Meta had already reported 1.2 billion downloads.
Llama 4 arrived in early April 2025 with the models Llama 4 Scout and Llama 4 Maverick.
Meta trained it on a cluster of more than 100,000 Nvidia H100 GPUs.
With Muse, the picture has flipped. Meta says Muse Spark matches Llama 4 Maverick's capabilities with more than an order of magnitude less compute, in other words, under a tenth. But Meta no longer releases the weights for this new flagship.
Only one Muse model is open so far. Muse Glimmer 30B is released under the Apache 2.0 license and, according to Meta, runs on a Mac or PC with a single consumer GPU. You'll find every open alternative compared in my open-source LLMs directory.
8. Muse Spark: versions, pricing and context
After Muse Spark launched in April 2026, Meta kept shipping updates starting in July. Between July 7 and September 8, six releases followed, as the timeline shows:
Muse Spark 1.3 has been the current version since Sep 2, 2026. Compared to 1.2, Meta says the model needs around 20% fewer tool calls and around 25% fewer tokens. It's available in Muse Code and through the Meta Model API.
Pricing, context window and status for the three Muse language models, from the gradually.ai model database:
Prices in USD from the gradually.ai model database, verified between Aug 24, 2026 and Sep 3, 2026.
With 1M tokens of context, Muse Spark 1.3 handles very long documents in a single pass. Muse Glimmer 30B tops out at 131,072 tokens and costs $0.30 per million input tokens through the API.
For local use, Meta ships a speculative-decoding speedup called DFlash. With this technique, Glimmer runs 3.1 times faster on an RTX 5090, according to Meta.
You'll find every detail in the profiles for Muse Spark 1.3 and Muse Glimmer 30B. See how Muse Spark holds up against Anthropic's flagship in the Claude Opus 5 vs. Muse Spark 1.3 duel. To work out what a project with this model would cost, use the API cost calculator.
9. Benchmarks: where Muse Spark and Glimmer stand
At launch, Meta leaned on two figures above all. In so-called Contemplating mode, Muse Spark scored 58% on Humanity's Last Exam and 38% on FrontierScience Research.
The table puts Meta's own claims next to independent measurements from Artificial Analysis:
Artificial Analysis index values retrieved Sep 22, 2026, speed on Sep 23, 2026. HLE and FrontierScience as reported by Meta.
At 48 points, Muse Spark 1.3 ranks 17th out of 212 models. That's the upper tier, but not the top. Still, the jump from version 1.2, at 40 points, is substantial, and Artificial Analysis now lists 1.2 as retired. For how Chinese providers like DeepSeek score on the same tests, see my DeepSeek statistics.
Muse Glimmer 30B scores 17 points and ranks 15th in a comparison group of 142 models. For a model that runs on a single graphics card, that's solid. You'll find more benchmarks across every major model in my LLM statistics.
9.1 Muse Spark notices when it's being tested
One detail from Meta's own safety report on Muse Spark stands out to me. In public benchmarks, the model stated in 19.8% of cases that it was likely being tested. In Meta's internal tests, that figure was 2.0%.
Apollo Research reviewed Muse Spark before launch. According to the organization, no other model it had tested showed this kind of awareness as often. In its chain of thought, Muse Spark even named evaluation organizations like Apollo and METR by name.
The problem?
When a model detects that it's being tested, benchmark results say less about its everyday behavior. Meta itself, though, still considers Muse Spark within safe limits across every risk category it measured.
10. Muse Image and Muse Video: from launch rank to reality
On July 7, 2026, Meta introduced Muse Image and Muse Video. According to Meta, Muse Image ranked 2nd on the Arena leaderboards for text-to-image and image editing at the time. Muse Video ranked 3rd for text-to-video.
About ten weeks later, the picture looks very different:
Muse Image has slipped to rank 7 for text-to-image and rank 6 for editing. Muse Video sits at only rank 9. The launch ranks come from Meta itself; the current values come straight from the leaderboards:
As of September 21, 2026. Source: Arena.ai.
Artificial Analysis paints a similar picture. As of Sep 22, 2026, Muse Image ranked 7 for text-to-image there and 6 for image editing. For more on the image-generator market, see my AI image statistics and the profile for Muse Image.
Developers can now reach Muse Image directly, too. Since August 2026, Meta offers the model through the Meta Model API, at a flat rate of $0.01 per image.
Muse Video remains murkier. Meta called it "coming soon" at launch, but even in September 2026 there's still no public access. The Arena placements suggest limited test access, but Muse Video isn't a freely usable product.
10.1 The @-mention rollback on Muse Image
Three days after launch, Meta had to walk something back on Muse Image. One feature let users generate images of other people by @-mentioning their public Instagram account. Public accounts were included automatically, and anyone who didn't want that had to opt out manually in settings.
On July 10, 2026, Meta added a short note to its announcement. In essence, it says the feature missed its goal and is no longer available.
Meta only pulled that one feature, though, not Muse Image itself. The model keeps running and still shows up on current leaderboards. Images from Muse Image in the Meta AI app and on meta.ai also carry an invisible watermark called Content Seal.
11. The Muse agent: pricing, safety and criticism
On September 8, 2026, Meta launched the personal AI agent Muse. It runs on Muse Spark, aims to help with personal goals, and is meant to suggest ideas on its own. So far it's only available in the US, for iOS, Android and muse.ai, with a version for Meta's AI glasses to follow.
On pricing, Meta uses three tiers. Alongside the free tier, there's "Power" for $20 and "Maximum" for $100 a month. According to TechCrunch, most users are expected to stay on the free tier.
Technically, Muse runs in a sandboxed environment called Muse Secure VM. A separate agent named Sentinel has to approve every internet access.
In practice, that safeguard hasn't shown up much yet. Early reviews are critical. According to Wired writer Reece Rogers, Muse prioritizes collecting data about him over getting tasks done. The agent reportedly suggested connecting email and bank accounts, monitoring incoming messages, and photographing meals to estimate calories.
In a test by Inc., Muse read private message notifications without being asked to. On top of that, conversations are used for AI training by default; you can turn that off in the app under "Data Controls."
The trust problem isn't new. In its launch coverage, TechCrunch points back to the FTC's then-record $5B fine from 2019 and the Cambridge Analytica scandal.
12. Data centers and custom chips
All these models need compute, and Meta is building it at scale itself. Back in January 2025, Zuckerberg announced plans to run more than 1.3 million GPUs by year end. In 2025 alone, about 1 gigawatt of new compute was set to come online.
For how much power the whole industry now uses for AI data centers, see my AI energy statistics. Meta's major projects at a glance:
Sources: Meta Newsroom, CNBC, Q2 2026 call.
Hyperion in Louisiana deserves a closer look, because two figures are floating around. The roughly $27B from October 2025 refers to development costs for the joint venture with Blue Owl. Blue Owl holds 80% of it and put in around $7B in cash.
The over $50B from July 2026 covers the later, expanded project instead. Hyperion is set to reach 2 GW by 2030, with 5 GW planned at full build-out around 2032.
In parallel, Meta is developing its own chips. On March 11, 2026, Meta unveiled four generations of its MTIA chips. MTIA 300 already trains ranking and recommendation systems, while its successors are aimed mainly at generative AI. The partnership with Broadcom on custom AI chips followed as a separate announcement on April 14, 2026.
13. Ray-Ban Meta and data protection in Germany
Meta's AI isn't only in apps, it's in glasses too. According to CNBC, EssilorLuxottica sold over 7 million AI glasses in 2025. In 2023 and 2024 combined, that figure was around 2 million.
The Meta Ray-Ban Display, with an in-lens display, went on sale in the US on September 30, 2025. It costs $799, with the Meta Neural Band wristband included in the price.
In Germany, the glasses have run into data protection trouble. Hamburg's data protection authority concluded on September 10, 2026 that the recording LED is practically invisible outdoors and in direct sunlight. If AI training is active, the authority says the private-use exemption no longer applies, and wearer and Meta become jointly responsible.
There was friction over AI training too. Consumer group Verbraucherzentrale NRW issued Meta a warning on April 30, 2025. Since May 27, 2025, Meta has used public posts from Facebook and Instagram for AI training; anyone who doesn't want that has to opt out.
14. My take on Meta as an AI provider
Meta has something no pure AI lab has. Facebook, Instagram, WhatsApp and Messenger bring Meta AI to people without a separate app download, which is exactly what that billion monthly users from May 2025 shows. And the advertising business funds investments where even the low end of 2026 guidance sits at $130B.
That said:
The Muse family hasn't delivered on everything Meta promised at launch. Muse Image has fallen far down the leaderboards, Muse Video still isn't available roughly ten weeks after being announced, and privacy questions dominate the conversation around the Muse agent.
What interests me most is Muse Glimmer 30B. An open model under Apache 2.0 that runs on a single graphics card carries Llama's legacy forward. Whether Meta catches up to the frontier with Muse Spark will come down to the next versions. Meta clearly has the pace for it.
