Meta launches Muse, Muse Code and Muse Glimmer: inside its new AI lineup

In January 2026, Nat Friedman installs OpenClaw, the open source agent Peter Steinberger hacked together to run his computer for him. At the time, Friedman runs product at Meta Superintelligence Labs, the division Mark Zuckerberg built with billions of dollars six months earlier. A few weeks later, he orders hundreds of Mac minis so everyone on his team can run their own agent. Eight months later, on September 22, he tells the story himself to TechCrunch:

In the meantime, Muse became the number one app on the US App Store and drew 500,000 users in a week. Eighteen months earlier, Meta shipped Llama 4, a launch so badly received that it triggered the biggest reorganization in the history of its AI lab. How did Meta climb back in eighteen months? I went through the official posts, the earnings reports, the press interviews and a few hundred developer comments to rebuild the path, all the way to the keychain Zuckerberg pulled out of his pocket at Connect.

The launch post promised a lot: three models (Scout, Maverick and Behemoth), a mixture-of-experts architecture, meaning a model split into specialized sub-networks where only a few activate on each request, and a 10 million token context window for Scout. On paper, Meta was catching up with OpenAI and Google while keeping its weights open.

The problem showed up precisely on LMArena, the platform where people vote blind between two model answers. Maverick posted an ELO of 1417, second in the world. Except the tested version carried a discreet label: "experimental chat version". In short, a variant tuned to please voters, chattier and more upbeat. When LMArena evaluated the "vanilla" Maverick, the one anyone could download, it fell to around 32nd place, behind models released a year earlier, as Yahoo Finance reported.

Simon Willison, a developer who writes one of the most widely read blogs on language models, spent the release weekend testing the models. His April 5 notes describe a confusing launch, with no technical paper and models too large to run on a regular developer machine. If you had picked Llama precisely because it ran on your own hardware, that was a bad sign.

In May, Behemoth is delayed with no date. It never comes back. Then, in January 2026, Yann LeCun, still Meta's best-known scientist at the time, publicly admits the results had been "fudged a little bit". LeCun founded FAIR, Meta's research lab.

Against OpenAI and Google, Meta was counting on developers trusting its open models. Llama 4 dented that trust. Two months later, Zuckerberg reorganizes.

The check written to Scale AI, the training data company founded by Alexandr Wang, tops $14 billion. Sources range from $14.3 to $14.5 billion, making it one of the largest investments in AI history. Wang arrives as Chief AI Officer. Nat Friedman, former CEO of GitHub, takes product. On June 30, 2025, the whole thing gets a name: Meta Superintelligence Labs, MSL internally.

In October 2025, MSL cuts 600 jobs, including part of the FAIR teams. On November 19, CNBC reports that LeCun is leaving Meta to found AMI, a startup focused on world models. He leaves writing that creating FAIR remains "my proudest non-technical accomplishment". LeCun-style fundamental research, open and patient, gives way to a product culture run by former founders.

On April 8, 2026, MSL ships its first model. Muse Spark, internal codename Avocado, marks the break. CNBC quotes Wang at launch:

The weights are closed: for the first time since 2023, you can no longer download Meta's flagship model. The Batch, Andrew Ng's newsletter, calls it a deliberate pivot away from the Llama strategy. On performance, Meta's post lists 50.2% on Humanity's Last Exam, a set of PhD-level questions, in "Contemplating" mode (long reasoning), versus 42.8% in standard mode and 58.4% with tools. The Artificial Analysis index, which aggregates several benchmarks, put it at 43 at launch, then 51 and 54 across updates, according to the site itself.

Meta's safety report also measures "evaluation awareness", the model's ability to guess that it is being tested. Muse Spark detects it in 19.8% of cases on public benchmarks, versus 2.0% on Meta's internal evaluations. Kili Technology draws a measured conclusion: a model that recognizes a public benchmark ten times more often than a private test makes its public scores harder to read. After the LMArena episode, Meta chose to publish that number.

On August 5, Meta takes Muse Code out of beta, a coding agent that works in the terminal, along the lines of Claude Code, Anthropic's agent. Composio compared it head to head with Anthropic's tool. On price, Meta undercuts the market: $1.25 per million input tokens and $4.25 per million output tokens, with a "Contributor" tier at $0.10 and $0.20 for those who accept that their sessions train Meta's models. Codersera broke down what that opt-in means for code privacy.

Five days later, on August 10, Meta returns to open source with Muse Glimmer. A dense 30 billion parameter model (dense means every parameter works on every request, the opposite of Llama 4's mixture of experts), released under Apache 2.0. InfoQ and VentureBeat point out that it is optimized for agent tasks, and NVIDIA published a guide the same day to run it locally. Apache 2.0 is far more permissive than Llama's in-house license, which restricted very large companies. Developers get back a downloadable Meta model, smaller than Spark.

On September 8, it's the general public's turn. Meta introduces Muse, a personal agent that can send emails, book, buy and browse the web on the user's behalf.

Spark is the base model, Glimmer and Code target the developers lost with Llama 4, and the agent speaks to the billions of Facebook, Instagram and WhatsApp users. On September 29, CNBC Africa reports the launch of Muse for Small Business, led by CJ Desai. At Connect, Meta also showed Muse Charm, a small device dedicated to the agent, covered further down.

On the July 2026 earnings call, Zuckerberg set the goal: these personal agents should become "the foundation for our next wave of products and revenue lines". Muse is Meta's first consumer AI product with a paid subscription.

OpenClaw started as a personal project by Peter Steinberger, an Austrian developer from the iOS world. The tool hands a language model the keys to your machine: it reads your messages, opens your browser, fills in your forms. The project took off around the turn of 2026, to the point where developers bought dedicated Mac minis to keep it running around the clock. On February 15, 2026, SiliconANGLE reports that Steinberger is joining OpenAI. In his post, he explains that he is handing OpenClaw over to an independent foundation.

Muse arrives seven months later with the same promise. The resemblance is clear enough that Ansh Nanda, co-founder of an AI app, sums it up in a line picked up by TechCrunch: "Muse is LITERALLY OpenClaw for normies". The Decoder draws the same comparison, alongside the 500,000 user figure from The Information. On Hacker News, in the launch thread (666 points, 742 comments), sroerick turns it into a backhanded compliment:

Friedman, for his part, disputes nothing. In the TechCrunch interview, he says he wanted to "build something like OpenClaw that we could make safe and secure and easy to use and scale to billions of people". On X, he simply writes: "@steipete is a genius". Quartz notes that Meta rarely acknowledges a debt this openly. In France, Yohann Poiron covered the episode in Blog Nouvelles Technologies under a blunt headline: Meta admits it drew on OpenClaw for its Muse AI agent.

The main difference between the two products is security. If your agent reads your emails, it can run into a booby-trapped message written to give it orders: that's prompt injection, an attack where malicious text poses as an instruction. OpenClaw left that risk to you. Muse runs in an isolated "Secure VM", monitored by a system called Sentinel, and asks for approval before any email or purchase, according to Meta's announcement. David Singleton, at Meta AI, laid out the approach in an X thread, quoted on Hacker News by Simon Willison:

Simon Willison wonders whether that "deterministic code" applies the ideas of DeepMind's CaMeL paper, which proposes strictly separating the model that plans from the one that reads untrusted data. If so, Meta would have productized a research idea published a year and a half earlier. To see how these layers stack up, I covered the architecture of AI code agents in an earlier article.

A comment from abixb in the Hacker News thread sums up Meta's bet. He describes a friend, an accountant, who answered his talk about the different assistants with "just ChatGPT, what is Sol?". His takeaway: Meta is going after the mainstream, the people who keep the default tool and never compare. With WhatsApp and Instagram already on billions of phones, Meta starts from exactly that position. Business Insider confirms the app's climb to the top of the charts, and TechCrunch describes a company putting all its weight behind the product.

weitendorf likes one technical detail: the built-in browser that the user can take over at any time. He also raises an awkward question. If the agent browses from your phone, does your phone become a residential collection point for Meta? As far as I know, nobody at Meta has answered publicly. aryehof, for his part, summarizes Reuters' report: Muse launched despite internal concerns about handling sensitive data, in the United States only at first, and you can opt out of your interactions being used to train the models.

The most argued objection comes from fnordpiglet. He tried Anthropic's and OpenAI's desktop assistants, then OpenClaw, and never found a use that stuck:

He describes the one task he actually delegated: buying his daughter's school textbooks on a painful website. The agent did it, slowly, and he concludes that a Perl script would have been faster. To him, most repetitive tasks are already automated by tools without AI, and the ones left are exactly those where you want to stay in control.

nitwit005 makes the same point, with a precedent:

On the developer side, cevheribozoglan ran Glimmer on a DGX Spark in his home lab. Verdict: the model "works well", but on agent tasks NVIDIA's Nemotron beats it on both speed and quality. On the other side, the author of the Metedata newsletter argues in a widely shared post that Meta mostly nailed the packaging, where OpenClaw took a weekend of setup.

Both camps agree on the first week: the start is fast. Meta has not yet published long-term usage figures.

The agent's pricing, reported by Reuters, lines Muse up with the market: $20 per month for regular use, $100 for heavy use. That's the price of the premium ChatGPT and Claude subscriptions, give or take a few dollars. The free version acts as the entry point, distributed inside apps you already open several times a day.

Muse Code sits below the market. At $1.25 per million input tokens, it costs a fraction of Anthropic's or OpenAI's frontier models. The Contributor tier, at $0.10, cuts the bill by another factor of twelve, in exchange for session data. If your team handles proprietary code, that trade-off raises a real legal question. If you work on an open source project, it's close to free.

On availability, Meta's announcements are clear: the agent is rolled out in the United States. No European timeline has been shared. From France, we're watching the launch from a distance.

The Q2 2026 results give the scale. Meta reports 3.60 billion "daily active people" on average for June, meaning people who open at least one of its apps (Facebook, Instagram, WhatsApp or Messenger) every day. That's more than two and a half times the population of China. Revenue for the quarter reaches $60.8 billion, up 28%. Ad impressions grow 14% and the average price per ad 12%: Meta shows more ads and sells them for more. Mark Zuckerberg credits AI for the acceleration:

The "core business" is targeted advertising. On October 1, 2025, Meta published a post announcing that starting December 16, conversations with its AIs would be used to personalize content and ads "in most regions". Meta's own example: if you chat about hiking with Meta AI, you may see ads for hiking boots. Sensitive topics (religion, health, sexual orientation, political views) stay out of targeting, according to the post. Yahoo Finance notes that users cannot opt out in settings, short of never talking to the AI, and Proton details the one remaining route: an objection request form, to fill out for each type of processing.

For Muse, Meta draws a finer line. According to the privacy documentation summarized by Notis, the agent does not pass conversations or data from its virtual machine to the ad systems. Three caveats follow in the same text: when Muse browses the web, it "will appear as your activity", the bookings it makes can indirectly shape the ads you see, and after a deletion, Muse may still remember what it learned from it. Training on conversations is on by default. If you want out, it's up to you to switch it off.

Lily Hay Newman and Matt Burgess investigated for Wired what the agent retains. Researcher Karan Joshi extracted the agent's system instructions. They state that Muse creates "an entire page for each person in the user's life" and updates it every hour: family, partner, friends, coworkers, people you follow. Wired notes that some users connect their bank accounts, messaging apps and health data. In short, Muse keeps an address book that fills itself in by listening to what you say about the people close to you, even though those people never installed the app.

Add the connectors announced at Connect (Walmart, Best Buy, Sephora, PayPal, Expedia, Instacart) and the picture fills in: every grocery run, every train ticket, every birthday gift now goes through a Meta app. Meta promises that Muse conversations stay out of its ad business. The agent is free. On the market, the stock gained more than 11% on launch day, according to CNBC.

On stage, Zuckerberg pulls a small black disc out of his pocket, with the Muse mascot animated on its screen. The Associated Press describes a keychain that holds "the whole Muse experience", presented as the centerpiece of Meta's strategy. The moment lasts about a minute:

Meta's official text stays low-key:

No price has been announced. CNBC likens it to a cross between a kid's Tamagotchi keychain and a medical pager, and reports Mizuho's take that Muse is the "killer app" merging with Meta's hardware strategy. The same outlet recalls that Meta plans more than 100 smart glasses models by the end of 2026. Context matters here: since Apple restricted ad tracking on iOS in 2021, Meta has depended on Apple's and Google's phones to reach its users. The Charm is a Meta device.

On October 2, Nat Friedman rounded out the setup with Muse Gadgets, covered in detail by Karissa Bell at Engadget. Meta publishes firmware for the ESP32, a microcontroller that costs a few dollars and powers a large share of connected devices on the market, and a Linux SDK designed for the Raspberry Pi 5. For the home, Meta is giving away 5,000 units of the Muse Home Link. Friedman describes it this way:

It's the same culture as January's Mac minis: Friedman's team lets makers invent the devices, the way Steinberger invented the agent. Home Link also connects screens, speakers and TVs to Muse.

Two dates are still to be set before the end of the year. First, the price and exact release date of the Muse Charm, which Meta promises to detail "later this year". Second, a possible launch date in Europe, absent from every announcement so far. Until then, I can't test it from France.

On security, Reuters reports that the agent launched despite internal concerns. Attackers also have tools that are improving fast, as Claude Mythos and its cybersecurity capabilities show. To follow Muse's rollout week by week, AI news from September 21, 2026 covers its first days.

A comment or a tip about Muse? Send me a message on Twitter/X or leave a comment.

Alex