1.2 billion downloads. That's how many times Meta's open Llama models had been downloaded by the end of April 2025.
For its Meta AI assistant, Meta reported more than 1 billion monthly active users in October 2025.
Then Muse showed up.
Since April 2026, Muse Spark has been Meta's proprietary language model, joined by Muse Wonder on October 12. The name also covers image and audio models, an announced video model, an open model, and a coding agent. Since September, even a personal AI agent is simply called "Muse." Keeping track of all that takes some doing.
Here are the key Meta AI statistics for 2026 on users, Muse, and Llama, separate from company revenue and investment. The three main points first:
- Meta reported more than 1B monthly active Meta AI users on its Q3 2025 call
- For 2026, Meta expects capital expenditure of $135-145B, up from $72.22B in 2025
- Muse has several models and products; Muse Image ranked ninth on Arena's Oct 6 text-to-image snapshot, while no public Muse Video access was confirmed by Oct 7
1. Meta AI at a glance
Meta develops AI models and products, but earns most of its revenue from advertising while investing in AI infrastructure. Here are the key figures:
As of Oct 28, 2026. Sources: Meta Investor Relations, Meta Newsroom.
- Meta Platforms (Facebook, Instagram, WhatsApp, Messenger)
- Meta Superintelligence Labs, founded June 30, 2025
- Alexandr Wang
- more than 1B monthly actives (Q3 call, Oct 29, 2025)
- Muse Wonder (Oct 12, 2026)
- Muse Glimmer 30B (Apache 2.0)
- 1.2B (April 2025)
- 3.64B average for September 2026 (+3% year over year)
- $135-145B (as of October 21, 2026)
- $64.60B (+26%)
- 66,775 (Sep 30, 2026)
In the third quarter of 2026, $63.06B of Meta's $64.60B in revenue came from advertising. Meta does not report how individual AI investments are funded.
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 products for consumers and businesses. The table lists the main offers; Muse for Small Business extends the personal agent rather than introducing a separate base model:
As of Oct 28, 2026. Sources: Meta Newsroom, Meta Model API, research.meta.ai.
- What it is
- Language model from Meta Superintelligence Labs, natively multimodal
- Launch
- Apr 8, 2026, currently version 1.3 (Sep 2, 2026)
- Access and status
- meta.ai, Meta AI app, Meta Model API, proprietary
- What it is
- Language models for personal and professional agentic tasks (Muse Wonder and Muse Wonder Full)
- Launch
- Oct 12, 2026
- Access and status
- Muse Wonder via Meta Model API and Muse Code, Muse Wonder Full for Muse Code subscribers; proprietary
- What it is
- Image generation and editing
- Launch
- Jul 7, 2026
- Access and status
- at launch in the Meta AI app, on meta.ai, in Instagram Stories (US) and WhatsApp (select countries)
- What it is
- Video generation
- Launch
- announced Jul 7, 2026
- Access and status
- not publicly released
- What it is
- Open language model with 29.6B parameters
- Launch
- Aug 10, 2026
- Access and status
- Apache 2.0, runs locally on one consumer GPU
- What it is
- Real-time audio model for transcription, speaker diarization and endpointing
- Launch
- Sep 1, 2026
- Access and status
- Meta Model API, Meta AI for Mac and Muse Code; 70+ training languages, 25 validated, 20+ speakers
- What it is
- Terminal coding agent
- Launch
- Aug 5, 2026
- Access and status
- Out of beta since Sep 24, 2026; uses Muse Spark
- What it is
- Personal AI agent powered by Muse Spark
- Launch
- Sep 8, 2026
- Access and status
- Rolled out in the US on Sep 8; described as available in the US and Canada on Sep 29
- What it is
- Business expansion of the personal Muse agent, not a separate base model
- Launch
- Sep 29, 2026
- Access and status
- Business accounts and connectors; approval before publishing, sending or spending
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.
On September 1, 2026, Meta also introduced Muse Voice Transcribe, Meta Superintelligence Labs' first real-time audio perception model. Meta lists streaming transcription, speaker diarization and endpointing among its tasks. It was trained on more than 70 languages, 25 of which Meta says it has extensively validated.
The model page lists more than 20 speakers and audio input exceeding one hour as supported scenarios. Meta lists the Model API, Meta AI for Mac and Muse Code as access points.
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. In Europe, the assistant first launched in 41 countries and 21 overseas territories, according to Meta's rollout announcement.
3. How many users does Meta AI have?
For its assistant, Meta reports only occasional milestones in monthly active users. For all its apps combined, the company publishes daily active people every quarter.
3.1 Monthly active Meta AI users
Meta announced successive usage milestones. Its product update reported more than 400 million on September 25, 2024; Mark Zuckerberg said nearly 500 million at Connect that same day. In December, it reported nearly 600 million.
On its Q4 call on January 29, 2025, Meta said it had more than 700 million monthly actives; in March, it reported that scale again. Zuckerberg said about 1 billion in May.
On its Q3 call on October 29, Meta reported more than 1 billion monthly actives.
As of October 31, 2026, Meta had not reported a newer absolute count of monthly active Meta AI users than the more than 1 billion it gave on its Q3 call on October 29, 2025. The Q2 2026 call described only a relative increase in daily interactions:
On its Q2 2026 call, Meta said the number of people interacting with its assistant each day had risen 60% since the Muse Spark rebuild.
Meta also said more than 1 million businesses used Business Agents weekly on WhatsApp and Messenger during Q2; at the time of the call, the company was also rolling them out on Instagram.
The standalone Meta AI app launched on April 29, 2025. The European rollout began on March 19, 2025, across 41 countries and 21 overseas territories, initially with text chat in six languages. Meta said at the time that the version launched in Europe had not been trained on EU user data.
3.2 Daily active people across Meta's apps
For comparison, all apps in the family together reached 3.64 billion people every day on average in September 2026. Meta calls this number Family Daily Active People (DAP) and counts each person once, even when they use several apps.
The figure dipped slightly in March 2026, which Meta attributed to internet disruptions in Iran and a restriction on WhatsApp in Russia.
gradually.aiFor how AI usage overall is developing, see my AI statistics.
4. Revenue, advertising and AI in the ad business
Meta uses AI in ad delivery, recommendations and ad creative production. In the third quarter of 2026, revenue rose to $64.60B.
Revenue grew 26% year over year, also at constant currency. Advertising alone grew 26%, according to Meta's Q3 results.
Meta sees its biggest lever in Advantage+, its automated campaign management. According to CFO Susan Li on the Q2 2026 call, 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.
Meta reported on its Q4 2024 call that more than 4 million advertisers used at least one generative AI ad creative tool, up from 1 million six months earlier. On the Q1 2026 call, the company reported more than 8 million advertisers.
For Q2 2026, Meta counted only small businesses rather than all advertisers and reported more than 9 million that used at least one AI creative tool.
On its Q2 2026 call, Meta attributed an 8.3% increase in ad clicks and a 15.7% increase in conversions on Facebook to its changes to ranking and user-understanding models. An early Instagram pilot with language models was associated with 1% more app-event conversions, according to Meta.
4.1 Revenue and net income since 2021
For the full year 2025, Meta's revenue hit $200.97B, up 22% year over year, with net income of $60.46B.
gradually.aiRevenue has risen every year since 2023 (it dipped slightly in 2022), while net income swung more and dipped slightly in 2025 despite higher sales.
4.2 Revenue by region in 2025
By customer address, 2025 company revenue split across regions like this:
gradually.aiThe figures cover total revenue, almost 98% of which comes from advertising ($196.2B of $201.0B). Europe includes Russia and Turkey, and rest of world covers Africa, Latin America and the Middle East.
5. Capex: Meta's investment spending is rising
Capex stands for company-wide capital expenditure, including spending on data centers, servers and networking gear. In 2024 it came to $39.23B, according to SEC filings.
For 2026, Meta guided to $135-145B in its Q3 results, more than triple. According to that report, $34.91B went into capital expenditure in the third quarter of 2026, after $31.08B in the second and $19.84B in the first. The figures include finance-lease principal.
gradually.aiMeta first forecast $60-65B in capex for 2025 in its Q4 2024 results; its FY2025 results later showed actual spending of $72.22B.
5.1 Capex guidance for 2026
The 2026 guidance changed in the Q4 2025 results, the Q1 2026 call, the Q2 2026 results and the Q3 2026 results:
The ranges include principal payments on finance leases. In the last two revisions, only the lower end moved up, first to $130B and then to $135B in the Q3 results, while the upper end stayed at $145B. Meta described each as a narrowed range.
5.2 Reality Labs operating losses
Reality Labs, the glasses and VR division, has posted an operating loss every year since 2020, and the loss kept growing through 2025:
gradually.aiFor 2026, Meta expects a loss of $18B to $19B, according to its Q3 results. For 2027, it points to roughly $15B as it shifts spending mainly toward AI.
Meta reported 75,472 employees on June 30, 2026 and, per its Q3 report, 66,775 on September 30, 2026, down 15% from a year earlier. On the call, CFO Susan Li put the drop at 12% from the second quarter. For comparison, check my OpenAI statistics.
5.3 Voting power of the largest shareholders
The ownership structure shows who decides on this spending. Meta has two share classes: Class B carries ten votes per share, Class A one. Mark Zuckerberg holds almost all Class B shares and so controls 60.8% of the voting power, although he owns only about 13.5% of the capital by calculation.
gradually.aiThe large fund managers BlackRock and FMR together hold more than 11% of the capital but only just over 5% of the votes.
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: Meta, SEC filings and dated media reports.
- June 30, 2025 (memo by Mark Zuckerberg)
- Alexandr Wang (Chief AI Officer)
- Meta said more than $14.3B; reports: about 49% non-voting and roughly $29B valuation
- about 600 roles in Meta's AI division, according to media reports (October 2025); no confirmed MSL headcount
- Muse Spark (April 8, 2026)
In June 2025, Meta agreed to a strategic investment in Scale AI. Reuters reported $14.3B for about 49% without voting rights and a valuation near $29B; Axios reported about $15B and a valuation above $29B.
Meta's SEC filing confirms a non-voting minority stake. Scale founder Alexandr Wang joined Meta as Chief AI Officer.
In October 2025, the Associated Press reported about 600 role cuts in Meta's AI division. Meta did not publish a verified MSL headcount after the reorganization. In November 2025, Yann LeCun announced plans to leave Meta at year-end and start a company focused on world-model research.
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. At LlamaCon in late April, Meta cited 1.2 billion downloads, according to TechCrunch; Meta’s LlamaCon recap confirms the billion-download milestone.
Llama 4 arrived in early April 2025 with the models Llama 4 Scout and Llama 4 Maverick.
On Meta's Q3 2024 call, Zuckerberg said Meta was training Llama 4 models on a cluster with more than 100,000 Nvidia H100 GPUs.
In its Muse Spark launch report, Meta said the model reached capabilities similar to Llama 4 Maverick while using more than ten times less compute in its internal comparison. Meta does not publish Muse Spark's weights.
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
Since Muse Spark launched in April 2026, Meta has announced several model versions, tools and agent features. The timeline records dated releases and product announcements:
Muse Spark 1.3 has been the latest released version since Sep 2, 2026. In Meta's own comparison with 1.2, it used about 20% fewer tool calls and about 25% fewer tokens. Muse Code left beta on September 24; Meta said the Meta Model API was generally available after its Connect update.
8.1 Context windows of the Muse models
Meta lists a 1 million token context window for Muse Spark 1.3. The chart shows the context window; release and status sit next to each model name:
gradually.ai8.2 API pricing for Muse Spark and Wonder
The current Meta Model API pricing has two tiers for Spark 1.3, while Muse Wonder comes in the Standard tier only. The chart shows the prices per million input, output and cached input tokens:
Meta says Standard prompts and completions are not used to train its models. In the discounted Contributor tier, users permit Meta to use prompts and completions to train future models.
Muse Voice Transcribe costs $0.18 per hour of audio processed; audio time is billed separately from token usage. Streaming and non-streaming transcription use the same rate. The audio service did not have a Contributor tier at launch.
For local use, Meta describes DFlash as a speculative-decoding method. Meta reports substantially faster generation on several tested devices at comparable output quality.
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 highlighted two scores. In so-called Contemplating mode, Muse Spark scored 58% on Humanity's Last Exam and 38% on FrontierScience Research, according to Meta's launch report.
The chart shows the Artificial Analysis Intelligence Index v5, retrieved on October 28, 2026:
Artificial Analysis gives Muse Wonder 43 points and rank 22 of 51, and Muse Spark 1.3 36 points and rank 29. Artificial Analysis no longer treats Spark 1.3 as current. Index v5 lists no value for Spark 1.2.
Glimmer scores 6, below the median of 13 for open models in its size class. For how Chinese providers score on the same tests, see my DeepSeek statistics.
You'll find more benchmarks in my LLM statistics.
9.1 Muse Spark notices when it's being tested
Meta's safety report measured how often Muse Spark suspected that certain benchmark prompts were tests. It did so in 19.8% of responses to cases classified as public benchmarks, compared with 2.0% in Meta's internal test set.
Meta's report also summarizes an Apollo Research evaluation before launch. Apollo observed unusually frequent test recognition.
The problem?
If a model recognizes test situations, that can affect benchmark measurements. Meta's report nevertheless concluded that Muse Spark stayed within its release thresholds across the risk categories it tested.
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.
An Arena snapshot from September 21 showed different ranks. The chart shows the comparison, the table below the latest snapshot for each board:
On October 6, Muse Image ranked ninth on Arena's text-to-image board and eighth for single-image editing. For text-to-video, September 21 remained the latest retrievable snapshot, with Muse Video at rank 9.
Leaderboard snapshots: image models Oct 6, 2026; text-to-video Sep 21, 2026. Votes are crowdsourced samples and do not prove public product access. Source: Arena.ai.
- Leaderboard
- Text-to-image (overall)
- Elo
- 1,274 (±5)
- Rank
- 9
- Votes
- 41,538
- Leaderboard
- Single-image editing
- Elo
- 1,403 (±4)
- Rank
- 8
- Votes
- 142,644
- Leaderboard
- Text-to-video
- Elo
- 1,456 (±15)
- Rank
- 9
- Votes
- 2,188
For more on the image-generator market, see my AI image statistics and the profile for Muse Image.
Developers can also access Muse Image directly. Meta offers the model through the Meta Model API. The pricing page lists $0.01 for each image successfully generated and returned.
Muse Video remains less clear. Meta called it "coming soon" in July; no public access was documented by the cutoff.
10.1 The @-mention rollback on Muse Image
At launch, Meta announced a feature that could use Instagram profiles as references in image prompts.
On July 10, 2026, Meta updated its announcement, saying the @-mention feature had missed its goal and was no longer available.
Meta withdrew only that feature, not Muse Image. The model keeps running and still appears on current leaderboards. According to Meta's launch report, images created with Muse Image in the Meta AI app and on meta.ai carry Content Seal, an invisible watermark.
11. The Muse agent: pricing, safety and criticism
On September 8, 2026, Meta introduced its personal AI agent Muse. It uses Muse Spark and, according to Meta, runs in its own virtual environment. The launch announcement described a US rollout.
At Connect on September 24, Meta announced plans to bring the agent to its AI glasses. In its September 29 announcement, Meta described Muse as available in the US and Canada.
Axios reported at launch that Muse had a free tier, "Power" at $20 and "Maximum" at $100 per month.
Meta describes Muse Secure VM as a dedicated virtual machine for the agent and user data. A control system called Sentinel is meant to monitor external actions; Meta says Muse asks for approval before sensitive steps such as sending an email or making a purchase.
Early tests raised questions about the balance between usefulness and data access. Wired writer Reece Rogers described suggestions to connect email and financial accounts, monitor messages and photograph meals.
On October 3, Wired reported that extracted Muse instructions described creating a page for each person in a user's life and updating it hourly. Meta told Wired that Muse draws on public information and what a user chooses to share.
Inc. columnist Jason Aten reported that Muse synced private message notifications on his Mac without his knowingly approving access. Meta disputed the account, saying message syncing and macOS access permissions are optional and that the product had explained syncing unclearly.
Meta says users can opt out in Data Controls of having their interactions used to train its AI models, as described in its launch announcement.
The trust problem isn't new. In 2019, the FTC imposed a $5B penalty for Facebook's violations of a prior privacy order and deceptive claims about users' ability to control their data.
At Connect on September 24, Meta announced real-time voice conversations and video chats with the agent, plus plans to bring Muse to its AI glasses. On September 28, its Enterprise Platform announcement described plans to bring Muse, Meta Business Agent, Muse API and Muse Code to businesses.
A separate September 29 announcement, Muse for Small Business, added business accounts and third-party connections. Meta says nothing will be published, sent or spent without approval.
At Connect, Meta also announced Muse Charm, a pocket-sized device intended to provide access to the agent through a real-time voice model. Meta said it was aiming for a holiday 2026 launch; it had not published technical specifications or a price by the cutoff.
TechCrunch reported that Muse reached the top of the US App Store on September 18 and remained there on September 25; the app was then available in the US and Canada.
Download estimates on September 25 varied: Sensor Tower estimated more than 3.4 million, Apptopia 4.3 million and Appfigures about 2.3 million. On September 30, 9to5Mac reported that a Sensor Tower projection had passed 5 million downloads and that Muse had held the top US App Store position for 12 days.
12. Data centers and custom chips
All these models need compute. On the Q4 2024 call, Zuckerberg set a target of more than 1.3 million GPUs in operation by year-end and nearly 1 gigawatt of compute capacity in 2025.
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:
At Hyperion in Louisiana, the roughly $27B described in Meta's October 2025 joint-venture announcement refers to total development costs for the venture. Blue Owl-managed funds own 80% and Meta 20%.
Meta said the funds made an approximately $7B cash contribution to the venture and that Meta received a separate $3B distribution.
In July 2026, Meta expanded its plan to 5 GW of capacity and put the project investment at over $50B. Meta did not give a completion year for the 5 GW stage.
Meta says Prometheus is designed for 1 GW of capacity once complete but did not give an online date on its project page. For its El Paso campus, announced with BlackRock in July 2026, Meta expects 1 GW of capacity to come online from 2028.
Meta is also developing its own chips. On March 11, 2026, it described four MTIA generations. Meta said MTIA 300 was already used for ranking and recommendation systems; later generations are also intended for generative AI. Meta announced a separate agreement with Broadcom on custom AI chips 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. In its FY2025 results, EssilorLuxottica reported more than 7 million AI glasses sold across its portfolio in 2025, as The Information reported on February 11, 2026. The earlier estimate of about 2 million covered Ray-Ban Meta glasses sold from their 2023 launch through the end of 2024.
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. It is the most expensive model in the lineup:
gradually.aiOn September 10, 2026, Hamburg's data protection authority published a report limited to the first-generation Ray-Ban Meta Wayfarer.
It concluded that recording other people outside one's close family may generally be unlawful because the LED does not provide enough transparency for informed consent. The authority also found that there was usually no legal basis for training AI on such recordings.
Meta said it began using adult EU users' public posts and interactions with Meta AI to train its own models on May 27, 2025, with an objection process.
It says public posts by EU accounts under 18 are not used, and private messages are excluded unless a user or chat participant shares them with Meta AI.
14. How popular is Meta?
Meta scores 79 out of 100 in the AI vendor popularity index, rank 5 of 38. The value combines search demand for the vendor and its AI brands, usage via OpenRouter and Hugging Face and mentions in tech media.
In the EU, Meta is at rank 6 of 36 (70 points), in Germany at rank 8 of 36 (64 points). Use the switch above the chart to change between world, EU and Germany.
Vendor popularity index
0 to 100, higher is more popular
As of Nov 7, 2026 · Source: gradually.ai popularity index
15. How popular are Meta's models?
Meta's most popular model is SAM 3 with 68 out of 100 (rank 7 of 785 AI models with a popularity score). In total, 24 Meta models have a score; the chart highlights them.
In the EU, Muse Spark 1.3 is at rank 16 of 613 (69 points), in Germany, Muse Spark 1.3 is at rank 15 of 692 (66 points). Use the switch above the chart to change between world, EU and Germany.
Popularity of Meta models
Popularity index from 0 to 100, higher is more popular.
29 of 119 models
Source: gradually.ai popularity index
16. How good are Meta's models?
In the gradually.ai overall index, which combines independent benchmarks into one value, Muse Wonder Full is Meta's strongest model: rank 8 of 92 with 118.3 points. 3 Meta models are ranked in the index. 100 points is the average of a fixed reference group; 15 points more means a model wins about three of four benchmark comparisons.
Overall index (Gradually Meta-Score v4)
20 of 92 models
4 models in this view are measured at a lower reasoning effort than max in the overall index. Dashed column: lower reasoning effort than the top.
17. How often do Meta's models hallucinate?
Artificial Analysis uses AA-Omniscience to measure how often a model gives a wrong answer to a knowledge question instead of admitting it does not know. Meta's best result comes from Muse Wonder Full: 18% of its answers that are not correct are wrong guesses (rank 4 of 62, rank 1 hallucinates least). The median of all measured models is 39%.
5 Meta models have a measurement. Each model counts with its best measured setting. The rate describes reliability, not safety in the sense of jailbreaks or harmful content.
Hallucination rate
Share of wrong answers among all answers that are not correct, lower is better. Best measured setting per model.
21 of 62 models
Measured by Artificial Analysis (AA-Omniscience)
18. How fast are Meta's models?
Muse Wonder is the fastest with 124 output tokens per second. The median of all measured models is 87 tokens per second. Artificial Analysis has measured 4 Meta models.
Speed
24 of 100 models
Output tokens per second, measured by Artificial Analysis
19. How expensive are Meta's models?
Via the API, Meta's cheapest model, Llama 4 Scout, costs $0.15 per 1M tokens. The most expensive is Muse Wonder at $8.00. Both are blended prices of three parts input to one part output, as in the LLM price index.
API prices
Blended price in US dollars per 1M tokens, 3 parts input to 1 part output
Source: gradually.ai LLM price index
20. My take on Meta as an AI provider
Meta places its assistant inside products many people already use. Its latest reported absolute scale was more than 1 billion monthly active Meta AI users, reported on its Q3 2025 earnings call. Advertising remained central, at $63.06B of $64.60B in Q3 2026 revenue. The company's capex guidance for 2026 is $135-145B.
That said:
The Muse products are still developing. Muse Video had no public access documented by October 7; Muse Image ranked ninth on Arena's text-to-image board on October 6. The US app-store ranking dates from September. Questions about data access and safety remain, especially because media reports of individual tests and Meta's description do not fully agree.
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.
Frequently asked questions about Meta AI
On its Q3 2025 call, Meta reported more than 1 billion monthly active Meta AI users. Through October 31, 2026, I could not verify a newer absolute count in Meta’s public disclosures, including the Q3 call on October 21.
Muse Spark is the language model from Meta Superintelligence Labs and the first model in the Muse family. The first version arrived on April 8, 2026, and the current version, Muse Spark 1.3, followed on Sep 2, 2026. Meta has not released weights for the Muse Spark 1.3 version offered through Muse Code and the Meta Model API. On October 15, 2026, Meta separately released weights for Muse Spark Open, a distinct open-weight variant, under the Muse Model License with limited commercial terms.
Meta AI is the assistant in Facebook, Instagram, WhatsApp, Messenger, the Meta AI app, and meta.ai. Muse Spark is a model in the Muse family that powers Meta AI. Muse is a separate personal AI agent powered by Muse Spark. According to Meta’s September 29 announcement and its October 21 Q3 call, the agent is available in the US and Canada and has not launched globally.
Yes. Meta’s July 10, 2026 update withdrew the Instagram @-mention feature for image prompts, not Muse Image itself. Meta describes Muse Image as available in Meta AI and lists it in the Meta Model API at $0.01 per image.
Meta announced Muse Video in July 2026 as “coming soon” to creators and Meta AI. On September 24, Meta also announced real-time video generation with Muse. I could not verify public access to Muse Video in Meta’s announcements reviewed through October 31, 2026.
Meta reported company-wide capital expenditures of $72.22 billion in its 2025 results, including principal payments on finance leases. Its July 29 Q2 results put 2026 capex at $130 billion to $145 billion; on October 21, Meta narrowed that outlook to $135 billion to $145 billion in its Q3 results. Both ranges include finance-lease principal.
Meta reported more than 650 million Llama downloads in December 2024. Its March 18, 2025 announcement put the total at more than one billion. At LlamaCon on April 29, 2025, Meta cited 1.2 billion downloads, according to TechCrunch; Meta’s own LlamaCon recap repeats the billion-download milestone. Meta’s Open Source AI page lists “1.2B+” downloads without a date. I found no later dated total in Meta announcements checked through October 31, 2026, including the Q3 call on October 21.
Meta AI began rolling out in Europe on March 19, 2025, across 41 European countries and 21 overseas territories, initially with text chat in six European languages. Its European training notice says it uses adults’ public posts and interactions with Meta AI to train and improve its models, and provides an objection process for public data. The separate Muse agent is available in the US and Canada, according to Meta’s September 29 announcement.
