Meta Has Two Open-Weight Families Now, and a Paid Tier
Say out loud which thing you mean. That one habit is worth more than any feature list here, because Meta now ships under two different licenses and reuses names that collide. Meta runs one hosted proprietary track and two open-weight families. The Meta Model API serves Muse Spark, whose documented current version is 1.3. Muse Glimmer is a 30B open model published under Apache 2.0, which means the 700 million monthly-active-user condition in the Llama license does not reach it. Meta also launched Meta One, a global paid subscription, on September 15, 2026, and a standalone personal AI agent called Muse on September 8. Choosing the track is still the first decision. There are now four of them.
One hosted track, two open licenses, one paid tier
The old split was “proprietary Muse for consumers, hosted Llama for developers.” It stopped holding in July, and the framing that replaced it has already changed again. Meta’s own endpoint serves Muse and names no Llama model. Open weights arrive from two families under two different licenses. Llama 4 sits under the Llama Community License, with its 700 million monthly-active-user condition. Muse Glimmer sits under Apache 2.0, with no such condition. And above both sits Meta One, a paid subscription that went global on September 15, 2026.
Muse Spark 1.3
Built by Meta Superintelligence Labs. It powers the Meta AI assistant, meta.ai and the new standalone Muse agent. Meta never announced 1.2 or 1.3. The last dated version announcement is 1.1, from July 9, 2026, so a version check made against Meta’s AI blog will still report 1.1.
· Meta’s model page describes 1.3 as trained for agentic workflows and optimized for competitive coding performance.
- Version 1.3, not 1.1
- The only model named on the Meta Model API
- No downloadable weights
Muse Glimmer
A 30B open model Meta describes as built for always-on local agents, running on a single GPU, tuned for tool use, long tasks and failure recovery. It is published under Apache 2.0, not the Llama Community License. That is the most buyer-relevant licensing change in the period. Meta now has two open-weight families, not one.
- Apache 2.0 removes the 700 million user condition
- Single-GPU deployment target
- Serving documented for vLLM
Llama 4
Scout and Maverick are natively multimodal open-weight models. They are still available through Meta’s developer domain and partners, and still governed by the Llama Community License. They are no longer the newest open weights Meta has published.
- Deployment control
- Text and image understanding
- License conditions apply, including a user-count threshold
Meta One
A subscription announced on September 15, 2026 and available globally. Nine named tiers across single apps, consumer bundles and Creator and Business bundles. Meta reports more than 50 features launched and 15 million subscriptions and trials to date, and states that Meta AI will still be free for everyday use.
- The first real Meta AI business SKU
- Media generation, including video, is a subscriber benefit
- Global on announcement
The hosted Llama API
The widely repeated July 6, 2026 shutdown date has no first-party citation left. Meta’s own endpoint still names no Llama model, which is consistent with a shutdown. But consistent isn’t sourced. So don’t put that date in a client document without checking it yourself.
· Meta’s deprecation document for the hosted Llama API returns 404 as of September 18, 2026, and its host redirects to a new domain.
- The supporting source was observed returning 404 on September 18, 2026
- Meta’s own endpoint still names no Llama model
- Treat the shutdown date as unverified
The open-source resumption happened, and it landed off the AI blog
Meta’s AI blog really has published nothing newer than July 27, 2026, a run of 53 days to September 18. The silence is real, and it says nothing about what Meta shipped. Meta’s publishing moved. Consumer, safety and business announcements now land on about.fb.com/news, and model pages and a developer blog now live on developer.meta.com, with API documentation on ai.developer.meta.com. Four significant items landed away from the AI blog, and they split cleanly in two. Two are datable from first-party posts: Meta One, a global paid subscription, on September 15, 2026, and Muse, a standalone AI agent, on September 8, 2026. Two cannot be dated at all. Muse Glimmer, an Apache 2.0 open-weight model, carries no first-party release date. And the move from Muse Spark 1.1 to 1.3 is documented only on an undated model page. So neither of those two can be placed on a timeline at all. Anyone monitoring Meta has to watch the newsroom and both developer domains as well as the AI blog. And has to treat an undated developer page as a fact without a date rather than as a recent release.
Muse Glimmer Open weight, Apache 2.0
The Apache 2.0 license is the buyer headline. None of the Llama license conditions travel with this model. What is missing is the paperwork a procurement team asks for. Meta exposes no publication date, no context window, no exact model ID and no weights download location on any page that renders to a fetch. So confirm all four in a browser before you commit.
· Meta’s model page: an open 30B model for always-on local agents that runs on a single GPU, is Apache 2.0 licensed, and is tuned for tool use, long tasks and failure recovery.
· A developer blog post covers building with it. The Model API documentation covers serving it with vLLM for production-grade throughput behind OpenAI-compatible endpoints.
- Two open-weight families now, not one
- No 700 million user gate
- Release date, context window, model ID and weights location all unestablished
- Attributed to Meta’s developer pages, which carry no date
Muse Spark 1.3 Current
Meta’s model page titles the current release Muse Spark 1.3 and describes it as trained for agentic workflows and optimized for competitive coding performance. The 1.1 release from July 9, 2026 is the last version with a dated first-party announcement. 1.3 is documented on an undated model page.
· The July 9 description of a 1M-token context that the model actively manages, remembering actions, retrieving earlier work and compacting without losing steps, was written about 1.1. It has not been restated for 1.3.
- Meta claims agentic-workflow training and coding optimization
- Vendor claims: evaluate on your own workload
- Powers Meta AI, meta.ai and the standalone Muse agent
- No dated announcement for 1.3 could be found
Meta Model API Status unresolved
Meta’s own developer endpoint serves Muse Spark through an OpenAI-compatible interface and carries no Llama models. Its status label is truly unresolved, and that’s not a hedge. So do not restate the preview label as a checked fact. And don’t quote a tier or a price, because neither one can be read.
· The July 9, 2026 post called it a public preview.
· The current product page metadata says only that you can build with Meta’s AI models and access Muse Spark, with no preview label. The page body does not render to a fetch, and Meta’s pricing path returns 404.
- OpenAI-compatible package eases migration
- No downloadable Muse Spark weights
- Preview versus general availability: unresolved
- No readable pricing at all
Meta One September 15, 2026
A global paid subscription. Meta says plans are available globally, reports more than 50 features launched and 15 million subscriptions and trials to date, and states that Meta AI will still be free for everyday use. Subscriber benefits include more media generation with Meta AI, covering image creation and editing and video generation.
· Nine named tiers: Instagram Plus at $3.99 a month, WhatsApp Plus at $2.99, Facebook Plus at $3.99, Core at $7.99, Premium at $19.99. Then Creator and Business bundles, at Essential from $14.99, Advanced from $49.99, Expert from $149 and Max from $499.
- The Creator and Business bundles are the first real Meta AI business SKU
- Prices are Meta’s published starting points, not quotes
- Per-tier AI usage limits are not readable on Meta’s plans page
Muse, the standalone agent September 8, 2026
Meta launched a personal AI agent product named Muse. It is rolling out in the US on iOS, Android and muse.ai, and is stated as coming soon to AI glasses. Meta presents Muse as a distinct product rather than a rename of Meta AI. Whether it replaces, wraps or sits beside Meta AI in the long run is not documented anywhere first-party, and both brands are live at once.
· Meta describes it as powered by Muse Spark, its most capable model to date, built for real-world agentic work. Meta says it is free for most of what people need, with subscription plans for people who want to do more.
- A product name, not only a model-family name
- US only at launch, iOS, Android and muse.ai
- Relationship to Meta AI undocumented
Meta Business Agent June 3, 2026
Meta’s enterprise-facing agent offering, announced on the newsroom rather than on the AI blog. Meta is expanding its Business Agent to businesses of all sizes globally, across WhatsApp, Messenger and now Instagram, and reports more than one million businesses already using it. Getting started is free, and Meta stated that in the coming months businesses would access the agent through paid subscription offerings. The Meta One Creator and Business bundles launched three months later look like that paywall arriving. But Meta does not connect the two in either post, so treat the link as interpretation rather than as Meta’s statement.
- Free to start, paid subscriptions foreshadowed
- WhatsApp, Messenger and Instagram
- The paywall connection is interpretation, not vendor fact
Agentic Meta AI July 24
Meta announced assistant capabilities that act rather than only answer: daily briefings, recurring tasks, web research and slide generation, stated as powered by Muse Spark 1.1. No dated post since revisits the market scope, so on September 18, 2026 the July 24 post remains the only statement. The standalone Muse agent announced on September 8 covers similar ground under a different brand, which muddies the question of where agentic behavior actually ships rather than settling it.
- Rolled out that day in select markets
- WhatsApp support “in the coming weeks”, never since dated
- Scope unconfirmed rather than confirmed static
Muse Image Shipped
Meta’s image generation model, combining prompt reasoning, web context and multiple visual references inside the Meta AI experience. Shipped doesn’t mean everywhere, though. One cross-guide development is worth knowing. Muse Image arrived inside Runway on August 26, 2026 as the muse_image API model. So Meta’s image model is reachable through a competitor’s platform as well as Meta’s own apps.
· Meta’s July 7 post puts it in the Meta AI app and on meta.ai, in Instagram Stories in the US and on WhatsApp in limited countries, with Facebook listed as coming soon.
- Generate and edit
- Blend multiple photo references
- Text, QR codes and presets
- Also in Runway as muse_image since August 26, 2026
Muse Video Partly overtaken
Announced alongside Muse Image on July 7, 2026 as coming soon to creators and Meta AI, and never given an availability date under that name. The September 15 Meta One post lists video generation among subscriber benefits, so video generation is now a shipped and paywalled consumer capability. What Meta doesn’t do is name the model. Whether the Meta One video feature is Muse Video is unresolved. So do not write the brand name onto a shipped capability, and do not tell a client that Meta has no video generation.
- Video generation ships behind Meta One
- Meta does not name the model powering it
- No standalone Muse Video availability date has been published
Muse Code Known gap
A product page for Muse Code exists on Meta’s developer domain. Its contents can’t be read: the body does not render to a fetch, like every other page on that host. It is recorded here as a known gap rather than described from a page nobody has read. If a coding model matters to the decision, this is the page to open in a browser first.
- Existence confirmed, contents unknown
- Not described anywhere else first-party
- Highest-value item for a manual check after Muse Glimmer
The Zuckerberg letter August 10, 2026
Mark Zuckerberg published “The Future is for Everyone,” a positioning letter rather than a product announcement. That promise has since been kept, and it was kept off the AI blog where the letter itself appeared: Muse Glimmer is an open-weight model published under Apache 2.0. The letter’s other promises, a capable personal agent for everyone and a fully private mode Meta itself cannot see into, remain undelivered.
· The line that mattered, verbatim: “Now that Meta Superintelligence Labs are up and running, we will resume releasing some open source models soon.”
- The open source promise is now met
- The private-mode promise is not
- Quote as direction, not as a capability matrix entry
Llama 4 Scout Open weight
A 17B-active, 109B-total model with 16 experts, native multimodality and a supported 10M context window. Still governed by the Llama Community License.
- Efficient deployment target
- Fits on one H100 with Int4 quantization according to Meta
- Long-context and image grounding
Llama 4 Maverick Open weight
A 17B-active, 400B-total model with 128 experts, for stronger general multimodal, reasoning and coding work. Still governed by the Llama Community License.
- Higher serving requirement than Scout
- General-purpose open model
- Available from Meta and partners
Llama Guard 4
Meta’s safety model accompanies the Llama 4 ecosystem for classification and protection workflows. Nothing equivalent has been documented for Muse Glimmer.
- Part of a broader safety system
- Requires task-specific evaluation
- Does not replace application controls
Behemoth Not released
Meta described the Llama 4 teacher model as still training in the original announcement, and nothing since has changed that. Do not include it as an available model.
- No current downloadable model
- No procurement assumption
- Update only from a new official release
| Model | Access and license | Active / total parameters | Context | Primary use |
|---|---|---|---|---|
| Muse Spark 1.3 | Meta products and the Meta Model API; proprietary; status label unresolved | Not publicly disclosed | 1M and actively self-managed, stated for 1.1 and not restated for 1.3 | Meta AI assistant, the Muse agent, agentic tasks and product experiences |
| Muse Glimmer | Open weights under Apache 2.0; download location not established | 30B total; active count not disclosed | Not published | Always-on local agents, tool use, long tasks and failure recovery |
| Muse Image | Meta AI and selected Meta apps; proprietary | Not publicly disclosed | Visual conversation context | Image generation and editing |
| Muse Video | Announced July 7, 2026; not confirmed shipped under that name | Not publicly disclosed | Not publicly disclosed | Video generation; an unnamed video capability ships inside Meta One |
| Llama 4 Scout | Download and partners under the Llama Community License; no Meta-hosted API | 17B / 109B | 10M supported | Efficient long-context multimodal deployment |
| Llama 4 Maverick | Download and partners under the Llama Community License; no Meta-hosted API | 17B / 400B | 1M listed in Meta’s launch materials | Stronger open-weight generalist |
Meta puts intelligence where social and personal context already lives, and now sells it
Muse Spark is built for Meta’s distribution advantage. Billions of users, messaging, public content, cameras, shopping, social graphs and wearable devices. Two things changed in September 2026. Meta launched a standalone personal AI agent called Muse, which collides head-on with Meta’s own model-family name. And it launched Meta One, which puts a price on part of the assistant experience for the first time.
Muse, the personal AI agent
Launched September 8, 2026, rolling out in the US on iOS, Android and muse.ai, with AI glasses stated as coming soon. Meta describes it as powered by Muse Spark and built for real-world agentic work. It is free for most of what people need, with subscription plans above that. Meta presents it as a separate product from Meta AI, not a rename.
Meta One
Global from September 15, 2026. Nine named tiers, from $2.99 a month for WhatsApp Plus to $499 a month for the Max business bundle. Meta says Meta AI will still be free for everyday use, so the subscription buys extra features and capacity rather than access to the assistant itself.
Meta Business Agent
Announced June 3, 2026 and expanded to businesses of all sizes globally, on WhatsApp, Messenger and now Instagram, with more than one million businesses using it. Free to start, with paid subscription access foreshadowed for the coming months.
Briefings and recurring tasks
Announced July 24 and powered by Muse Spark 1.1: daily briefings, recurring tasks, web research and slide generation, in select markets with WhatsApp in the coming weeks. Meta has not revisited that scope, so as of September 18, 2026 it is unchanged on the record and unconfirmed in practice.
Parental alerts for teen distress
From July 16, Meta AI can alert parents when a teen’s conversation shows self-harm signals. Live in the US, UK, Australia and Canada, with global availability stated for year end. Meta added parental supervision on Threads for teens in the Asia-Pacific region on September 15, 2026.
Meta AI app and web
Rich conversation, personal context, web information and direct creation in one consumer surface, now running beside the separate Muse agent app.
WhatsApp and Messenger
Assistant access inside conversations where plans, questions and shared decisions already appear, and where the Business Agent meets customers.
Instagram, Facebook and Threads
Context from public posts, creators, groups and social content can shape search and recommendations.
AI glasses
Voice, camera and live visual assistance bring Meta AI into the physical moment. The Muse agent is stated as coming soon to glasses, and Meta Connect on September 23 and 24 is where any hardware change will land.
AI Mode in Facebook
Answers can be grounded in public perspectives from groups, Reels and other Meta content.
Marketplace plus web
Meta AI can combine local Marketplace listings with wider web products and map context.
Muse Image and video
Generate, edit and share visuals directly into stories, chats and feeds. Higher media generation limits, including video, sit behind a Meta One subscription.
Image generation grounded in the social and personal world
Output quality isn’t the whole proposition here. Muse Image can reason over the brief, combine several references and use web context inside the Meta AI experience. Two cautions now travel with it. One launch capability was withdrawn eight days after shipping. And as of September 15, 2026 the higher media generation allowances sit behind a paid Meta One plan. So check the current feature set and the plan boundary before you design a workflow around either.
Prompt reasoning
Muse Spark helps plan layout, retrieve relevant context and coordinate visual references before image creation.
- Multi-step generation
- Useful for informational visuals
- Review inferred facts
Multi-photo blending
Combine selfies, vacation images, pets, products and other references into a single creation.
- Check identity consistency
- Respect consent
- Avoid confidential uploads
Text and QR generation
Meta says Muse Image can render clean text and working QR codes for practical visual outputs.
- Test every QR code
- Proofread all text
- Do not trust generated regulatory copy
Presets
Suggested prompts make common transformations and styles reachable without writing a detailed brief.
- Useful for adoption
- Can create generic sameness
- Apply brand judgment
Sketch-based edits
Circle, sketch or annotate directly on the image to describe changes while keeping conversational context.
- Precise local direction
- Iterative editing
- Inspect unintended changes elsewhere
Withdrawn: profile mentions
The launch feature that let users pull public Instagram account content into a generated image was removed on July 15.
· Meta’s update reads: “We’ve heard the feedback that this feature missed the mark, so it’s no longer available.”
- Eight days from ship to withdrawal
- Do not document it as available
- Re-check any workflow built on it
Now partly paid
The September 15, 2026 Meta One post lists more media generation with Meta AI among subscriber benefits, covering image creation and editing and video generation. Free-tier limits and per-tier allowances are not published in a form anyone can read: Meta’s plans page does not render its body to a fetch. So assume a volume workflow needs a paid plan, and confirm the specific limit in the app.
- Media volume is a subscription lever
- Per-tier limits unverified
- Budget before promising throughput
Room redesign shopping
Use a room photo and ask for redesigns grounded in real products from the web or Marketplace.
- Visualize before buying
- Verify dimensions and availability
- Separate concept from specification
Advertiser path
Meta announced Muse Image support for Advantage+ creative, extending the model into paid-media production.
- Brand and claim review
- Test across placements
- Track model-generated asset provenance
Create three visual territories for this campaign using only the product images I upload. Explain the strategic idea before generating each route. Do not invent product features or customer claims.
Redesign this room for a hybrid-working professional. Keep the exact room geometry, use products that are currently available, and list the dimensions I must verify before purchase.
Create a 4:5 infographic from these verified facts. Keep text concise and legible. After generating, transcribe every word in the image so I can proofread it.
Keep composition, people and lighting unchanged. Remove only the marked object and rebuild the background naturally. Do not alter faces, clothing or product details.
Two open-weight families, two very different licenses
Open weights matter because you can download, host and customize the model. What changed in this window is that Meta now publishes two open families under two licenses. Llama 4 Scout and Maverick carry the Llama Community License and its 700 million monthly-active-user condition. Muse Glimmer carries Apache 2.0 and no such condition. Neither one is served by Meta. The serving decision belongs to you, or to a third-party host.
| Dimension | Muse Glimmer | Llama 4 Scout | Llama 4 Maverick |
|---|---|---|---|
| License | Apache 2.0 | Llama Community License | Llama Community License |
| User-count condition | None | 700 million monthly active users, then a license request to Meta | Same condition |
| Parameters | 30B total; active count not disclosed | 17B active / 109B total | 17B active / 400B total |
| Experts | Not published | 16 | 128 |
| Supported context | Not published | 10M | 1M in launch materials |
| Deployment positioning | Single GPU, per Meta’s model page | Single H100 GPU with Int4 quantization | Single H100 host, meaning one multi-GPU server rather than one GPU |
| Best fit | Always-on local and edge agents | Efficient extreme-context work | Stronger general open-weight work |
The license is the headline
Apache 2.0 on Muse Glimmer removes the single condition that dominates every Llama licensing conversation. No user threshold. No request to Meta. No discretionary grant. So if a client stalled on the Llama license, Muse Glimmer is the conversation to reopen. Don’t generalize the permissiveness, though. Llama 4 hasn’t been relicensed, and nothing suggests it will be.
- Two licenses, two risk profiles
- Check which model your vendor actually ships
- Record the license with the artifact
Serving Muse Glimmer
Meta’s Model API documentation covers serving Muse Glimmer with vLLM for production-grade throughput behind OpenAI-compatible endpoints. That is a familiar path for anyone already running open models, and it makes an A/B against a hosted endpoint easy to set up. The documentation page carries no date and its host renders nothing to an automated fetch, so treat the detail as attributed rather than cited.
- vLLM path documented by Meta
- OpenAI-compatible endpoints
- Confirm the doc in a browser
Long-context analysis
Scout’s supported 10M context can carry codebase, archive and multi-document workloads that are otherwise split across requests. Muse Glimmer publishes no context figure, so it cannot be compared here.
- Test retrieval at real lengths
- Build coverage checks
- Do not assume equal attention
Multimodal understanding
Scout and Maverick accept text and image inputs for grounding, visual question answering and application-specific workflows. Meta does not describe Muse Glimmer as multimodal.
- Use exact visual tasks
- Evaluate localization and detail
- Add application guardrails
Fine-tuning and customization
Open weights let an organization adapt behavior, domain language and task performance beyond prompting. Under Apache 2.0 the redistribution question is much simpler than under the Llama license.
- Curate training data
- Prevent regression
- Maintain evaluation baselines
Llama Stack and Llama Guard 4
Meta provides standardized building blocks for inference, evaluation, tools and application development, plus a companion safeguard model for content and risk classification. No equivalent safety companion has been documented for Muse Glimmer.
- One safety layer, not the whole system
- Tune policy to the use case
- Monitor false positives and misses
Two domains to repoint, not one
The Llama web estate has moved twice over. Every Llama link in a client’s procurement pack now points at one of two new hosts, and one old documentation path returns 404 outright.
· www.llama.com issues a 302 to developer.meta.com/ai/, and www.llama.com/llama4/license/ redirects the same way to developer.meta.com/ai/llama4/license/. Both re-confirmed on September 18, 2026.
· Separately, llama.developer.meta.com now 302-redirects to ai.developer.meta.com, which serves the Model API documentation.
- llama.com goes to developer.meta.com/ai/
- llama.developer.meta.com goes to ai.developer.meta.com
- The Llama API deprecation doc path is now a 404
- Repoint bookmarks, citations and contract annexes
Open models beyond the chat families
Meta’s open-model story is wider than Llama and Glimmer. On July 21 Meta described SAM 3 and DINOv3, both open source, deployed in SYNAPS-I, a Lawrence Berkeley led Department of Energy multi-lab project under the Genesis Mission. Meta’s research index now names Muse Image, Muse Spark, Muse Glimmer, SAM 3 and open-source AI models together. That corroborates Muse Glimmer as a first-class model line rather than an experiment.
- Vision models, not chat models
- Fine-tuned across 300 A100 GPUs
- Evidence of open models in serious scientific work
The open-weight advantage appears only if the organization can operate it
A downloadable model can improve control, customization and portability. It can also create a permanent model-serving, security and evaluation job. And with two Meta open families now in play under different licenses, the artifact you pin has to record which one you actually deployed.
What you cannot pin yet on Muse Glimmer
A pinned artifact needs a model ID, a weights location, a context configuration and a release date. Meta publishes none of the four on a page that renders to a fetch. So until someone opens the developer pages in a browser and records them, a Muse Glimmer deployment cannot be documented to the standard the Llama artifacts already meet.
- Model ID unknown
- Weights download location unknown
- Context window unpublished
- Release date not establishable first-party
Where the hosted option now sits
Meta’s own endpoint serves Muse Spark, not Llama and not Glimmer. So a hosted-versus-self-hosted comparison inside Meta’s estate is a comparison between different models, not different deployments of the same one. And neither the endpoint’s status label nor its pricing can be read, so the cost side of that comparison cannot be completed from public pages.
- Hosted means Muse Spark only
- No readable price
- Get a quote before modeling total cost
Cloud and edge partners
Llama is available through model hosts, clouds and edge ecosystems. Muse Glimmer’s partner availability is not documented on any Meta page that renders to a fetch. So ask any vendor claiming to serve it exactly which weights they hold, and under which license.
- Avoid provider-specific lock-in
- Check quantization and context differences
- Document the deployed artifact
Eleven habits for keeping Meta’s tracks and names straight
Name the product, the model, the license and the access path precisely. That is still the behavior that matters most, and it is harder than it was a month ago. Meta reused the word Muse for a consumer product, and it now publishes open weights under two different licenses.
Say which Muse you mean
On its own the word Muse means two things, the model family and the agent app. Four names remove the ambiguity. Muse Spark 1.3 is the hosted model. Muse Glimmer is the open model. Muse is the consumer agent app. Meta AI is the in-app assistant, which is not branded Muse. Never write only “Muse.”
Name the license, not just “open”
Muse Glimmer is Apache 2.0. Llama 4 is the Llama Community License, with a user-count condition. Open weight is not one thing.
Use Muse Spark for distribution
Choose Meta’s hosted track when social, messaging, camera, voice or creator context is the advantage.
Use open weights for control
Choose Glimmer or Llama when hosting, customization or data location is the requirement. And budget for the hosting Meta does not provide.
Do not say public preview
The Meta Model API carried that label on July 9, 2026, and the current product page neither repeats nor withdraws it. Say the status is unconfirmed, and quote no tier and no price: the pricing path returns 404.
Price the subscription
Meta One runs from $2.99 to $499 a month across nine tiers. Meta AI is still free for everyday use, so state which side of that line the client’s workflow falls on.
Test long-context coverage
A 10M window needs retrieval and reasoning evaluation at the real workload size. And Muse Glimmer publishes no window at all.
Treat social grounding as biased
Public Meta content reveals perspectives, not a representative population.
Review before direct sharing
Generation-to-feed convenience needs a deliberate pause before publication.
Operate the open model
Pin, secure, evaluate and maintain every deployed artifact, and record which license it came under.
Re-verify after Connect
Meta Connect runs September 23 and 24, 2026. Anything here about glasses, hardware or the assistant needs re-checking on September 25.
Reach, personal context, paid tiers and two open licenses create different risk profiles
Meta’s consumer products raise privacy, representation and publication questions. Its open weights push safety, security and lifecycle responsibility toward whoever deploys them, and now do so under two different sets of terms. And its documentation surface is largely unreadable to an automated check, which is a risk of its own.
The name Muse now means two things
Since September 8, 2026 the word Muse on its own means two things. One, Meta’s model family: Muse Spark, Muse Glimmer, Muse Image, Muse Video and Muse Code. Two, a standalone consumer agent at muse.ai, rolling out in the US on iOS and Android and stated as coming soon to AI glasses. Meta has not documented whether the agent replaces, wraps or sits beside Meta AI, and both brands are live at once. So any document that says “Muse” without qualification will be read two ways.
Llama license: the 700 million user clause, quotable at last
Open weight is not the same as unrestricted open source, and this is the clause that decides it. In plain words: if your product had more than 700 million monthly active users in the month before the Llama 4 release date, you have to ask Meta for a license, and Meta can refuse. Here is the text. Section 2 of the Llama 4 Community License Agreement reads: “If, on the Llama 4 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.” Muse Glimmer, under Apache 2.0, carries no equivalent condition.
· Sourcing caveat, stated plainly: this text was recovered on September 18, 2026 from Meta’s official GitHub mirror of the llama-models repository, not from the canonical license URL, which still cannot be read.
Why the canonical license page still cannot be read
The failure is characterized precisely here so nobody repeats the work. Nine pages across developer.meta.com and ai.developer.meta.com behave the same way, so this is a site-wide client-side rendering wall, not a block aimed at the license. Automated requests to every Meta host were separately refused at the proxy. The fix is a human with a browser and about thirty seconds.
· www.llama.com/llama4/license/ returns a 302 to developer.meta.com/ai/llama4/license/. That target returns head metadata only: a title reading “Llama 4 Community License Agreement,” a canonical URL and Open Graph tags, with a completely empty body. Retrying with the _fb_noscript=1 parameter returns an identical empty body.
Apache 2.0 is the exception, not a policy change
One Meta open model is now Apache 2.0. That doesn’t relicense Llama 4, and Meta has published nothing suggesting it will. So if a client says “Meta went permissive,” correct it. Meta published a permissive model beside a conditionally licensed family.
Muse Glimmer’s open questions
Confirmed to exist, with Apache 2.0 and 30B the only hard figures. No first-party release date can be obtained, so it cannot be placed on a timeline at all. Context window, exact model ID and weights download location are all unpublished on any page that renders. So don’t scope a deployment against it until those four are read in a browser.
Meta Model API status and price
The July 9, 2026 post called it a public preview. The current product page metadata omits any preview label, and the body does not render. The pricing path returns 404. Both the availability label and the cost are unknown, so commit to neither in a proposal.
Meta One cost and limits
Nine tiers from $2.99 to $499 a month, global from September 15, 2026. Meta states Meta AI stays free for everyday use. But it does not publish a readable definition of everyday use, or per-tier AI limits: the plans page renders an empty body. So a client planning volume media generation should establish the ceiling in the product before buying.
The hosted Llama API claim is now unsourced
Meta’s deprecation document for the July 6, 2026 shutdown returns 404, and its host redirects to a new domain. The shutdown is still plausible, and Meta’s endpoint still names no Llama model. But plausible and cited are different things. The date has no first-party citation today and is carried as an unsourced working assumption. So if an architecture decision depends on the date, verify it before the decision, not after.
Two domains to repoint
llama.com now redirects to developer.meta.com/ai/, and llama.developer.meta.com now redirects to ai.developer.meta.com. Those are two different hosts, and at least one old documentation path underneath the second returns 404 rather than redirecting. So every Llama reference in a client’s procurement pack, contract annex or internal wiki needs repointing. And a redirect isn’t a guarantee the specific page survived.
Staged rollout
The July 24 agentic capabilities started in select markets, with WhatsApp promised in the coming weeks, and nothing dated since updates that. The September 8 Muse agent launched in the US only. Both launch scopes are on the record and can be stated as such. What is unconfirmed is whether either has widened since, because no dated first-party page revisits them. So date the scope you quote and do not assume it still holds. Check the surface the client actually uses.
Features can be withdrawn
Muse Image shipped with public-profile mentions on July 7 and lost them on July 15. A shipped consumer AI feature can disappear in eight days, so don’t build a dependency on one.
Public-content bias
Groups, posts, Reels and profiles provide useful context without representing all users or reality.
Minors and duty of care
Parental alerts for teen distress reached the US, UK, Australia and Canada in July, with global coverage stated for year end. Threads added parental supervision for teens in the Asia-Pacific region on September 15, 2026. Coverage is uneven and expanding region by region, so check the specific market rather than the headline.
Direct sharing
The shortest path from generation to publication increases accidental or insufficiently reviewed output.
Intention is not capability
The August 10 letter promised to resume open source releases, and that promise has now been kept. Its other promises have not: an exceptionally capable personal agent for everyone, and a fully private mode where even Meta cannot see or grant access to your information. Neither one exists in a shipped Meta product today. So quote the letter as direction. Don’t enter it in a capability matrix.
Model operations
Self-hosting transfers serving, logging, patching, abuse controls and evaluation to the organization, for either open family.
Roadmap speculation
Don’t publish codenames, projected dates or unreleased models as product facts. No Llama 4.1, 4.5 or 5 exists on any first-party source. Muse Code has a product page whose contents cannot be read, which is a gap, not a product. Video generation ships inside Meta One, but Meta does not name the model, so do not call it Muse Video.
Connect will reset some of these sections
Meta Connect 2026 runs September 23 and 24, 2026, five days after the September 18, 2026 verification stamp. Meta describes two days on AI, wearable technology and immersive experiences. Hardware, glasses and assistant statements here are pre-Connect by construction. So re-verify on September 25, 2026 before reusing them.
First-party evidence behind this guide
Three dated first-party articles anchor the newest material: Meta One on September 15, 2026, the Muse agent on September 8, 2026, and Meta Business Agent on June 3, 2026. All three sit on Meta’s newsroom rather than its AI blog. One document that used to be citable no longer is. The Llama API deprecation page at llama.developer.meta.com returns 404 as of September 18, 2026. That took with it the only citation for the July 6, 2026 shutdown date, which is therefore carried as an explicitly unsourced working assumption. Several load-bearing facts here rest on pages that expose no date and render an empty body to an automated fetch, so they are attributed inline rather than cited. Confirm each one directly in a browser before it enters a client document. Source stamps say when each page was last opened. A fetch that didn’t happen is not a check.
· Attributed inline rather than cited: Muse Glimmer’s 30B size, Apache 2.0 license and single-GPU positioning, plus the developer blog post about building with it. Muse Spark 1.3 as the current version. The Meta Model API product page. The vLLM serving documentation at ai.developer.meta.com. Muse Code’s existence. The Meta Connect dates on meta.com. The Meta One plans page. And Meta’s research index.
· The verbatim 700 million user clause in Watch Outs was recovered from Meta’s official GitHub mirror of the llama-models repository. That is Meta’s own code channel but not a dated first-party article, so it is named there and deliberately not listed below.
AI Mindset