The Reach Machine That Also Ships Open Weights
Check the model ID before anything else. Meta sells two versions of Muse Spark 1.3. The plain ID, muse-spark-1.3, costs $1.25 per million input tokens and $4.25 output, and Meta says it isn’t used to improve its products. The contributor ID, muse-spark-1.3-contributor, costs $0.10 and $0.20, and that discount buys Meta permission to train future models on your content. Same model. Very different deal. And Spark isn’t the top of the line any more. Muse Wonder shipped October 12, 2026 at $4 in and $20 out, with a 1M-token context. Meta Enterprise Platform still has no price. And Meta’s own pages call the Model API both generally available and public preview.
One hosted track, two open licenses, one paid tier, one enterprise division
Five tracks. Pick one before you pick a model. First, hosted. Muse Wonder and Muse Spark run on the Meta Model API, with per-token prices in Meta’s docs. Muse Spark is on Oracle Cloud AI Platform too. Then two open-weight families under two licenses. Muse Glimmer is Apache 2.0, with weights on HuggingFace. Llama 4 carries the Llama Community License and its 700 million monthly-active-user clause. On either open path the hosting is yours or a partner’s, never Meta’s. Above that sits Meta One, a paid subscription live globally since September 15, 2026. And since September 28 there’s Meta Enterprise Platform, a division with a leader, a bundle and no price. Five doors, right?
Muse Wonder and Muse Wonder Full
Released October 12, 2026. Meta calls them its most capable models to date, built for personal and professional agentic tasks. Muse Wonder runs on the Model API and in Muse Code at $4 per million input tokens and $20 output, with a 1M-token context. Muse Wonder Full isn’t on the API at all. You get it only through a Muse Code subscription, on the $50 Power Usage or $200 Maximum Usage plan.
· Meta’s docs say Muse Wonder is in preview. It has no contributor mode.
muse-wonder-1.0on the APImuse-wonder-full-1.0in Muse Code only- Will soon power Muse, subscribers first
Muse Spark 1.3
Built by Meta Superintelligence Labs. It powers the Meta AI assistant, meta.ai and the standalone Muse agent. Meta announced 1.3 on September 2, 2026, available that day in Muse Code and the Model API. It’s still listed, with no deprecation label, and it now sits one rung below Muse Wonder. Version 1.2 is live too, and it’s what Muse Code runs by default.
· Two IDs per version. The plain one isn’t used to train Meta’s models. The contributor one is.
- $1.25 in, $4.25 out on the standard ID
- 1M context, listed as 1,048,576 tokens
- No downloadable Muse Spark weights
Muse Glimmer
A 30B open model Meta announced on August 10, 2026 for always-on local agents. Meta says it’s small enough for a Mac or PC with a single consumer GPU, and tuned for tool use, long tasks and failure recovery. Apache 2.0, not the Llama Community License. Text and image in, text out.
· It runs on your hardware, not on the Model API, so it has no Meta token price.
- Apache 2.0 removes the 700 million user condition
- 128K default context, longer supported
- Weights on HuggingFace, plus Ollama and llama.cpp packaging
Llama 4
Scout and Maverick are natively multimodal open-weight models, still governed by the Llama Community License. The license text and the download paths live on dev.meta.ai. They aren’t the newest open weights Meta has published, and Meta doesn’t host them for you.
- Deployment control
- Text and image understanding
- License conditions apply, including a user-count threshold
Meta Enterprise Platform
The biggest structural change of the fall, and the one with the least detail behind it. Announced September 28, 2026 as a new business division bundling the Muse agent, Meta Business Agent, the Muse API and Muse Code. Zuckerberg calls it the next major pillar of the business. CJ Desai, formerly chief executive of MongoDB, is Chief Enterprise Platform Officer. And that’s the whole of what Meta has published. No price, no tier, no date.
· Treat it as a stated intention with a named owner. Not a product you can specify.
- A division, not a SKU
- Senior hire signals real investment
- Zero commercial detail published
Muse for Small Business
Announced September 29, 2026. An agent that works in the background toward business goals and can connect to dozens of tools. Meta names Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, plus Facebook and Instagram business accounts, Pages and ad accounts. Nothing publishes, sends or spends without your approval.
· Meta says Muse is free for most of what people need, with subscription plans above that.
- Free for most needs, plans unpriced
- No separate country list for the small business product
- Approval gate on every outbound action
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.
- Creator and Business bundles are the consumer-side business SKU
- Media generation, including video, is a subscriber benefit
- Global on announcement
Muse Code
Meta’s terminal and CI coding agent. It launched in beta on August 5, 2026, left beta on August 31 and runs on macOS, Linux and Windows. It’s also the only place to get Muse Wonder Full. Four plans: $5, $15, $50 and $200 a month, with five-hour and weekly usage resets. No team plan is listed, only Contact Sales.
- Default model
muse-spark-1.2 - Muse Wonder Full from the $50 plan
- Part of the Meta Enterprise Platform bundle
The hosted Llama API
The widely repeated July 6, 2026 shutdown date still has no first-party citation. Third-party coverage attributes it to Meta. That’s a lead, not a source. What Meta does say, on an undated help page, is “Llama models aren’t run for you through Model API.” That’s a positive statement. It carries no date. So don’t put July 6 in a client document without checking it yourself.
· Meta’s old deprecation route, dev.meta.ai/docs/llama-api-deprecation, confirms no date either, so treat its current contents as unknown.
- Meta states Llama is not hosted on the Model API
- The shutdown date itself remains uncited
- Treat July 6, 2026 as a working assumption
Muse Wonder takes the top spot, and the prices are on the page
Meta’s biggest model release of the fall landed on October 12, 2026. Muse Wonder and Muse Wonder Full, which Meta calls its most capable models to date. That pushes Muse Spark 1.3 down a rung. And Meta now prints a full per-token price table for the Model API, contributor discounts included. So the cost side of a Meta decision is no longer a quote. Where did all this land? Not on the AI blog, which last posted on July 27. Meta’s model news now goes to research.meta.ai. Consumer and business news goes to about.fb.com. Prices and docs live on dev.meta.ai. And the enterprise division still has no price at all.
Muse Wonder and Muse Wonder Full October 12, 2026
Meta’s most capable models to date, by its own description, aimed at personal and professional agentic tasks. Muse Wonder, muse-wonder-1.0, is on the Model API and in Muse Code with a 1M-token context, at $4.00 per million input tokens, $0.20 cached and $20.00 output. Rate limits are 3,000 requests and 4,000,000 tokens per minute. Muse Wonder Full, muse-wonder-full-1.0, also has a 1M context, but it’s Muse Code subscription only.
· Meta names two features, Active Memory and Active Focus, and cites Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon. Those are vendor claims. Meta says Muse Wonder will soon power Muse, rolling out first to subscribers, and that a full report with more safety evaluations will be available soon.
- In preview per Meta’s docs
- No contributor mode, so no training discount
- No regions, clouds or open weights stated
- Safety report promised, not yet published
Meta Model API prices Published
Meta prints the price table on its Model API product page and in its pricing docs, per million tokens. Muse Wonder: $4.00 in, $0.20 cached, $20.00 out. Muse Spark 1.3, 1.2 and 1.1 on the standard ID: $1.25 in, $0.15 cached, $4.25 out. Muse Spark 1.3 and 1.2 on the contributor ID: $0.10 in, $0.002 cached, $0.20 out, in exchange for training rights. Web search grounding is $2.50 per 1,000 queries.
· Muse Image is $0.01 per image. Muse Voice Transcribe is $0.18 per hour. SAM 3.1 is $2.50 per 1,000 images and $0.20 per 1,000 video frames. The pricing docs state no free tier.
· Meta’s August 5, 2026 developer blog says Meta is beginning to accept requests for zero data retention.
- Read the ID before the price
- Contributor pricing trades your data for a discount
- Undated pages, so keep a dated copy
- Ask about zero data retention in writing
Meta Enterprise Platform September 28, 2026
Meta stood up an enterprise division and named a leader for it. The bundle is the Muse agent, Meta Business Agent, the Muse API and Muse Code. Zuckerberg’s framing is that this is the next major pillar of the business. CJ Desai, formerly chief executive of MongoDB, is Chief Enterprise Platform Officer.
· What’s missing is everything a buyer needs: no price, no tier, no availability date, no API specification. The article carries no update note, and no later newsroom post through October 14, 2026 mentions the platform. This is the biggest commercial gap in Meta’s AI estate, and the omission is Meta’s.
- A division with a named owner
- Bundles four existing products
- No commercial detail at all
- Do not put a number against it in a proposal
Muse for Small Business September 29, 2026
An agent that runs in the background against stated business goals and can connect to dozens of tools. Meta names fifteen: Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom. Plus Facebook and Instagram business accounts, Pages and ad accounts. Meta states that nothing publishes, sends or spends without your approval.
· Meta revised the article on October 8, 2026. Its pricing line is that Muse is free for most of what people need, with subscription plans. Its geography line, US and Canada, describes Muse the agent, not this product.
- Free for most needs
- Approval required before any outbound action
- Paid plans exist but are unpriced
- Country availability not stated for this product
Muse Glimmer August 10, 2026
Meta announced Muse Glimmer on August 10, 2026 as an open agentic model that runs on your device. Apache 2.0 is the buyer headline, and the paperwork a procurement team asks for exists.
· Weights at huggingface.co/meta-models/Muse-Glimmer-30B. The model card lists about 29.6B parameters, including a 1.8B vision encoder, in a dense 52-layer design derived from Muse Spark, and says it isn’t intended for users under 18. Context is 128K by default per Meta’s docs. It runs on your own hardware and isn’t called over the Model API. The name muse-glimmer is its Ollama tag. Runtimes named by Meta: Ollama, llama.cpp, ExecuTorch, LM Studio, Unsloth Desktop, vLLM and SGLang, plus Fireworks, Together.AI and OpenRouter.
- Two open-weight families now, not one
- No 700 million user gate
- Not a Model API model, so no Meta token price
- Dated, so it can go on a timeline
Muse Spark 1.3 September 2, 2026
Meta announced 1.3 on September 2, 2026: “available today in Muse Code and in Meta Model API.” Meta reports about 20% fewer tool calls and about 25% fewer tokens than 1.2, from an engineer comparison. Those are vendor claims. Meta’s docs list a 1,048,576-token context for 1.3, 1.2 and 1.1.
· The July 9 description of a context the model actively manages, remembering actions, retrieving earlier work and compacting without losing steps, was written about 1.1. Meta hasn’t restated that part for 1.3.
· Each of 1.3 and 1.2 has a standard ID and a contributor ID. Only the contributor ID lets Meta train on your content.
- Still listed, no deprecation label
- One rung below Muse Wonder
- Powers Meta AI, meta.ai and the Muse agent
- 1.2 is Muse Code’s default model
Muse Code Paid plans
Meta’s coding agent for the terminal and CI, with approvals and operating-system sandboxing. Beta on August 5, 2026, on macOS and Linux. Out of beta on August 31. Meta’s Connect recap added Windows.
· Plans: Everyday Usage $5 a month, with 10 to 50 prompts every five hours. High Usage $15, five times that. Power Usage $50, twenty times, and the first plan with Muse Wonder Full. Maximum Usage $200, eighty times, added October 12, 2026. Every plan gets Muse Spark, Muse Wonder and Muse Image.
· Install with curl -fsSL https://dev.meta.ai/install.sh | sh on macOS and Linux, or irm https://dev.meta.ai/install.ps1 | iex on Windows. Sign in by browser or API key, and use META_API_KEY in CI. Default model is muse-spark-1.2.
- No team plan listed, only Contact Sales
- macOS, Linux and Windows
- Five-hour and weekly usage resets
- Part of the Meta Enterprise Platform bundle
Muse, the standalone agent US, or US and Canada
Launched September 8, 2026 as a personal AI agent, powered by Muse Spark. Where is it available? Meta’s own newsroom gives two answers. The launch article, revised September 30, 2026, says Muse is rolling out in the US on iOS, Android and muse.ai. The Muse for Small Business article, revised October 8, 2026, calls Muse a personal AI agent available in the US and Canada. Treat Canada as unsettled until Meta says it in one place.
· New at Connect: Muse gets its own email address it can use to get things done. It also gets eighteen connectors. Retail: Walmart, Best Buy, American Eagle Outfitters, DICK’S Sporting Goods, Fanatics, Gap, Michael Kors, Sephora, Ulta and Wayfair. Payments: Shop Pay and PayPal. Travel and delivery: Expedia and Instacart. Work: Notion, Granola, GitHub and Box.
- Two Meta pages, two country lists
- Glasses access is coming months, not available
- Muse Wonder to power it soon, subscribers first
- Connectors reach commerce, payment and code systems
Muse Voice Transcribe and SAM 3.1 September 1 and March 27, 2026
Two models in Meta’s published list that aren’t chat models. Muse Voice Transcribe, muse-voice-transcribe-1.0, announced September 1, 2026 and available that day on the Model API, in Meta AI for Mac and in Muse Code. Meta says it was trained with 70 or more languages, 25 of them extensively verified. Price is $0.18 per hour, or $3.00 per 1,000 minutes.
· SAM 3.1, sam-3.1, announced March 27, 2026 as a drop-in replacement for SAM 3. On the Model API it costs $2.50 per 1,000 images and $0.20 per 1,000 video frames. Meta states no license for it on the model page or the announcement.
- Speech-to-text and segmentation, not chat
- No contributor mode for Voice Transcribe
- SAM 3.1 license unstated
Meta One September 15, 2026
A global paid subscription. 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. 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.
- Prices are Meta’s published starting points, not quotes
- Hearing Enhancement named as a plan benefit at Connect
- Per-tier AI usage limits aren’t in Meta’s announcement
Meta Business Agent June 3, 2026
Meta’s original business-facing agent, expanded to businesses of all sizes globally across WhatsApp, Messenger and Instagram, with more than one million businesses using it. Free to start, with paid subscriptions foreshadowed. Meta has since shipped two more things in that space: the Meta One Creator and Business bundles in September, and Muse for Small Business on September 29. The Business Agent is listed as a component of Meta Enterprise Platform.
- Free to start, paid access still unpriced
- WhatsApp, Messenger and Instagram
- Named inside the enterprise bundle
Agentic Meta AI July 24, 2026
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. The standalone Muse agent covers similar ground under a different brand, and Connect added connectors to Muse rather than to this surface. So where agentic behavior actually ships is still a question you answer per surface.
- Rolled out that day in select markets
- WhatsApp support in the coming weeks, never since dated
- Scope unconfirmed rather than static
Muse Image Shipped
Meta’s image generation model, combining prompt reasoning, web context and multiple visual references inside the Meta AI experience. It carries the model ID muse-image-1.0 and costs $0.01 per image on the Model API. Meta’s model page also lists it on fal, Runway and OpenRouter at the same price, so Meta’s image model is reachable through other platforms 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
- $0.01 per image on the Model API
Muse Video Still unnamed
Announced alongside Muse Image on July 7, 2026 as coming soon, and never given an availability date under that name. Video generation does ship, behind Meta One. But Meta attributes it to “Muse models” at family level and no video model appears in Meta’s published model list. One adjacent thing did get a name. On September 23, 2026 Meta described Muse Realtime Avatar, a system that generates expressive, interactive video avatars, paired with Muse Realtime Voice. It isn’t Muse Video and it isn’t in the model list.
- Video generation ships behind Meta One
- Muse Realtime Avatar is a named video-avatar system
- No video model ID exists in Meta’s model list
The Zuckerberg letter August 10, 2026
Mark Zuckerberg published “The Future is for Everyone,” a positioning letter rather than a product announcement. The open source promise in it has been kept: Muse Glimmer went out under Apache 2.0 the same day. The letter’s other promises are a capable personal agent for everyone and a fully private mode Meta itself can’t see into. Meta hasn’t announced either as shipped.
· 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 met
- No private mode announced as shipped
- 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. For Muse Glimmer, Meta points to a Usage Policy rather than a named companion safety model, so plan your own safety layer.
- 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. Don’t include it as an available model.
- No current downloadable model
- No procurement assumption
- Update only from a new official release
| Model | Model ID | Access and license | Parameters | Context | Primary use |
|---|---|---|---|---|---|
| Muse Wonder | muse-wonder-1.0 | Meta Model API and Muse Code; in preview; proprietary; no contributor mode | Not publicly disclosed | 1M (1,048,576 tokens) | Personal and professional agentic tasks |
| Muse Wonder Full | muse-wonder-full-1.0 | Muse Code subscription only, on the $50 and $200 plans; not on the Model API; proprietary | Not publicly disclosed | 1M | Meta’s most capable agentic tier |
| Muse Spark 1.3 | muse-spark-1.3 (not used for training); muse-spark-1.3-contributor (used to train Meta models) | Meta Model API; Oracle Cloud AI Platform; proprietary | Not publicly disclosed | 1M (1,048,576 tokens) | Meta AI assistant, the Muse agent, coding and agentic tasks |
| Muse Spark 1.2 | muse-spark-1.2; muse-spark-1.2-contributor (used to train Meta models) | Meta Model API; proprietary | Not publicly disclosed | 1M (1,048,576 tokens) | Default model for Muse Code |
| Muse Spark 1.1 | muse-spark-1.1 | Meta Model API; public preview for US developers per its model page; proprietary | Not publicly disclosed | 1M, actively self-managed per Meta | The model behind the July agentic Meta AI features |
| Muse Glimmer | No Model API ID; Ollama tag muse-glimmer; weights repo meta-models/Muse-Glimmer-30B | Open weights under Apache 2.0 plus a Meta Usage Policy; runs on your hardware | About 29.6B total, including a 1.8B vision encoder; dense | 128K by default, longer supported | Always-on local agents, tool use, long tasks and failure recovery |
| Muse Image | muse-image-1.0 | Meta AI apps; Meta Model API at $0.01 per image; also fal, Runway and OpenRouter; proprietary | Not publicly disclosed | Visual conversation context | Image generation and editing |
| Muse Voice Transcribe | muse-voice-transcribe-1.0 | Meta Model API at $0.18 per hour; Meta AI for Mac; Muse Code; proprietary | Not publicly disclosed | Not applicable | Streaming speech to text |
| SAM 3.1 | sam-3.1 | Meta Model API at $2.50 per 1,000 images; drop-in replacement for SAM 3; license unstated | Not publicly disclosed | Not applicable | Image and video segmentation |
| Muse Video | No model ID published | Video generation ships inside Meta One; no named model | Not publicly disclosed | Not publicly disclosed | Video generation, attributed to Muse models as a family |
| Llama 4 Scout | Not served by Meta | Download and partners under the Llama Community License; no Meta-hosted API | 17B active / 109B total | 10M supported | Efficient long-context multimodal deployment |
| Llama 4 Maverick | Not served by Meta | Download and partners under the Llama Community License; no Meta-hosted API | 17B active / 400B total | 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 four ways
Muse Spark is built for Meta’s distribution advantage. Billions of users. Messaging, public content, cameras, shopping, social graphs and wearables. And Muse Wonder is next: Meta says it will soon power Muse, subscribers first. Three things changed in September. The standalone Muse agent got its own email address and eighteen commerce and productivity connectors. Muse for Small Business shipped with a free core and an approval gate. And Meta Enterprise Platform appeared above all of it with no price attached. One thing is muddier than it looks. Is Muse the agent US only, or US and Canada? Meta’s newsroom says both, on two different pages.
Muse, the personal AI agent
Launched September 8, 2026. The launch article, revised September 30, says it’s rolling out in the US on iOS, Android and muse.ai. A newer Meta article, revised October 8, says Muse is available in the US and Canada. Meta describes it as powered by Muse Spark and built for real-world agentic work, free for most of what people need, with subscription plans above that. AI glasses access is still coming months.
What Muse can now reach
Connect added connectors for Walmart, Best Buy, American Eagle Outfitters, DICK’S Sporting Goods, Fanatics, Gap, Michael Kors, Sephora, Ulta, Wayfair, Shop Pay, PayPal, Expedia, Instacart, Notion, Granola, GitHub and Box. Meta also states that Muse will have its own email address it can use to get things done. That’s a new data and identity surface, not only a feature.
Muse for Small Business
September 29, 2026, revised October 8. Background agent work against stated goals, connecting to dozens of tools, fifteen of them named, including Shopify, Stripe, Slack, Notion, Canva, Figma, Asana, Klaviyo and Intuit QuickBooks. Nothing publishes, sends or spends without approval. Free for most needs, with subscription plans Meta hasn’t priced.
Meta Enterprise Platform
September 28, 2026. A division rather than a product, bundling the Muse agent, Meta Business Agent, the Muse API and Muse Code, under a Chief Enterprise Platform Officer hired from MongoDB. No pricing, tiers or dates. If a client asks what Meta sells an enterprise today, the real answer is the Model API, Muse Code plans and Meta One, with this on the horizon.
Meta One
Global since 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 named Hearing Enhancement for AI glasses as a plan benefit at Connect.
Meta Business Agent
Announced June 3, 2026 and expanded to businesses of all sizes globally, on WhatsApp, Messenger and Instagram, with more than one million businesses using it. Free to start. It’s named as a component of Meta Enterprise Platform, which is the clearest signal yet of where the paid version lands.
Briefings and recurring tasks
Announced July 24, 2026 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 hasn’t revisited that scope since, so it’s 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, 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. Meta added Forum on September 29, 2026, a standalone US app for Facebook Groups whose Ask feature surfaces existing group posts and comments.
AI glasses
Voice, camera and live visual assistance bring Meta AI into the physical moment, and Connect put prices and dates on the whole line. See the Hardware section. The Muse agent itself isn’t on glasses yet: Meta restated that as coming months on October 1, 2026.
Most of the glasses line is on sale, and the headset waits for spring 2027
Meta Adventurer goes on sale October 23, 2026 at $249. That’s the next hardware date that matters. Ray-Ban Meta Audio was due to ship October 13, and a Meta blog post on meta.com now says it’s on sale from $349. But that post carries no date, so nothing dated from Meta confirms units shipped on the 13th. Ray-Ban Meta Gen 3 has been on sale since September 23 from $449. Hearing Enhancement is still “later this year” in the US. And Meta VR Glasses, at $1,299.99, waits for spring 2027. A season isn’t a date. So plan around the dates and treat the seasons as hope.
| Device | Starting price | Availability | What it is |
|---|---|---|---|
| Meta VR Glasses | $1,299.99 | Spring 2027 | Roughly 100g two-part system: glasses plus a pocket compute puck on an optical tether |
| Ray-Ban Meta Audio | From $349 | Scheduled to ship October 13, 2026; an undated Meta post says on sale; India coming soon | Meta’s first audio-only glasses, 43g, no camera headline |
| Ray-Ban Meta (Gen 3) | From $449; India from INR 44,300 | Available September 23, 2026; India from October 1, 2026 | Camera AI glasses, third generation |
| Meta Adventurer | $249 | Pre-orders open; on sale October 23, 2026 | The cheapest device in the Meta Glasses line |
| Meta Ray-Ban Display | Not in the cited articles | US, Canada and UK from September 23, 2026; France, Italy and Germany scheduled October 13, 2026 | Display glasses with navigation and hologram calling in Early Access |
| LISA Collection and Starfire Ivory | From $399 | No date stated | Style editions of the Meta Glasses line |
| Hearing Enhancement | $149.99, HSA and FSA eligible, or with Meta One | US, later in 2026 | FDA-cleared hearing enhancement software for supported AI glasses |
| Muse Charm | Not announced | Product page live; no announced price or date | Meta’s first AI agent device, for talking to Muse |
Meta VR Glasses $1,299.99, spring 2027
The headline device, and the one furthest out. Roughly 100 grams in two parts: the glasses and a pocket compute puck joined by an optical tether. A 5K Infinite Display using micro-OLED at 37 pixels per degree. Qualcomm Snapdragon Reality Elite. Up to three hours of media playback with 45W fast charge. Meta calls it the first IMAX Enhanced certified VR device, with more than 75 games at launch and Xbox Cloud Gaming.
· Spring 2027 starts about five months after October 15, 2026, and it’s a season, not a date. Don’t schedule a deployment against it.
- Two-part design, not a standalone headset
- Three hours of playback is a real constraint for all-day use
- Price and window published; exact date not
Ray-Ban Meta Audio From $349
Meta’s first audio glasses, and the nearest thing to a mass-market entry point. 43 grams. Up to 12 hours of battery with another 48 from the case. 23 color and lens combinations across Clubmaster and Burbank styles. Prescription compatible. Meta’s September 23 article set shipping for October 13, 2026.
· A Meta blog post on meta.com titled “Ray-Ban Meta Audio, Our First Audio Glasses, Are Here” says it’s on sale from $349. The post carries no date. In India, Meta listed it as coming soon on October 1, 2026.
- Audio first, so a different privacy conversation than camera glasses
- Battery is the strongest spec in the line
- Check stock in your market before you promise it
Ray-Ban Meta Gen 3 From $449
The camera line, third generation, available from September 23, 2026. Up to nine hours of battery, one more than Gen 2. 27 color and lens combinations across three styles. A six-microphone array Meta says cuts more than 90 percent of background noise. A 12MP camera with 3K Ultra HD video. A customizable action button. Meta launched it in India on October 1, 2026, from INR 44,300.
- Camera and microphone policy applies here, not to the audio model
- Vendor claim on noise reduction: test it in your environment
- On sale now, so it’s a real purchase option
Meta Glasses line and markets October 23 for Adventurer
Meta Adventurer is the cheapest device in the line at $249, with pre-orders open and sales starting October 23, 2026. The LISA Collection, in Adventurer and Capri editions, starts at $399, as does the Kylie Jenner Starfire edition in new Ivory. New markets: Singapore, South Korea, Mexico and the UAE, with Brazil stated as coming soon. Meta’s stated target is more than 100 styles by year end across Ray-Ban, Oakley and Meta Glasses.
- $249 is the new floor for the line
- Four new markets, Brazil undated
- Style count is a vendor target, not a commitment
Meta Ray-Ban Display Europe from October 13
Online ordering opened September 23, 2026 in the US, Canada and UK, with retail in Vancouver, Toronto and London. Meta set October 13, 2026 for online ordering and retail in France, Italy and Germany, in Paris, Milan, Rome, Frankfurt, Hamburg and Munich. Meta’s supported-countries help page, which carries no date, lists the US, UK, Canada, France, Germany and Italy. Prescriptions come through EssilorLuxottica optical stores from October 13 in France and Italy, and later in the year in Germany.
· New features: cycling and public transit directions in the US and Canada, contextual walking directions that use landmarks rather than street names, Dolby Atmos spatial audio capture, an Explore app, enhanced calendar, mirrored phone notifications stated as coming soon, and Threads live feeds. A photorealistic hologram for WhatsApp video calls is Early Access in the US later this fall.
- Six markets on Meta’s undated help page
- Hologram calling is Early Access, not general
- German prescriptions come later than French and Italian
Hearing Enhancement $149.99 or with Meta One
FDA-cleared hearing enhancement software for supported AI glasses, aimed at adults with perceived mild to moderate hearing loss. Setup happens at home, with no clinic visit and no prescription. Meta says it’s launching in the US later this year. Price is $149.99, HSA and FSA eligible, or it comes with a Meta One subscription. No launch post appeared on Meta’s newsroom through October 14, 2026.
· This is the first time a Meta One subscription buys a regulated medical-adjacent capability, which changes who inside an organization needs to look at the plan.
- FDA cleared, which is a claim with a regulator behind it
- US only, no date beyond later in 2026
- Changes the Meta One buying conversation
Muse Charm No announced price or date
At Connect, Meta described a pocket-sized device built for talking and interacting with Muse, with real-time voice, and promised details later in 2026. Meta now has a product page for it, with regional versions for the UK, India, Australia, the UAE, Singapore, New Zealand and Ireland, describing it as its first AI agent device. No Meta announcement gives it a price, a date or a specification. Confirm the product page yourself before you quote anything from it.
- Product page exists
- No price or specification announced
- Do not count it as a 2026 purchase yet
Muse on glasses did not ship Coming months
The most repeated hardware expectation going into Connect was the Muse agent arriving on AI glasses: hands-free access, able to act on what the wearer is looking at without describing it. Meta restated it as coming months at Connect and again on October 1, 2026. It isn’t available. If a plan depends on a glasses-resident agent, that plan has no date behind it.
- Stated as coming months, twice
- Still not shipped
- The capability is described, not dated
Developer timing, if you are building Time-bound
Two Meta programs run on clocks worth knowing. The Meta Global AI Developer Hackathon has indie and startup tracks, a $1M prize pool and free Model API credits. The Meta VR Start Developer Competition opened September 24, 2026 and takes submissions until November 18, 2026, with $1M in total prizes across 20 cash prizes. The Wearables Device Access Toolkit left developer preview at 1.0, and web apps render on Meta Ray-Ban Display with no native app or SDK required.
- VR competition closes November 18, 2026
- Free Model API credits are a real cost offset for a pilot
- Wearables toolkit is at 1.0
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. It carries the model ID muse-image-1.0 and costs $0.01 per image on the Meta Model API, which makes a pilot cheap to price. Two cautions travel with it. One launch capability was withdrawn eight days after shipping. And since September 15, 2026 the higher media generation allowances inside Meta’s apps sit behind a paid Meta One plan. So check the current feature set, the plan boundary and the API route before you design a workflow around any of them.
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
- Review any workflow built on it
Partly paid in the apps, metered on the API
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. Meta’s announcement gives no free-tier limits or per-tier allowances. On the Model API the price is plain: $0.01 per image. So a volume workflow is easier to cost on the API than inside the apps.
- Media volume is a subscription lever in the apps
- $0.01 per image on the API
- 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 yourself. Meta 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. So same company, same word “open,” very different paperwork, right? And Meta serves neither one. It says so directly: Llama models aren’t run for you through the Model API, and Glimmer runs on your own hardware. Glimmer is fully specifiable: announced August 10, 2026, with a weights repository, a context window and an architecture on the record.
| Dimension | Muse Glimmer | Llama 4 Scout | Llama 4 Maverick |
|---|---|---|---|
| License | Apache 2.0, plus a Meta Usage Policy per the model page | Llama Community License | Llama Community License |
| User-count condition | None | 700 million monthly active users, then a license request to Meta | Same condition |
| Parameters | About 29.6B total, including a 1.8B vision encoder | 17B active / 109B total | 17B active / 400B total |
| Experts | None; dense, 52 layers | 16 | 128 |
| Supported context | 128K by default, longer supported | 10M | 1M in launch materials |
| Multimodal | Yes: text and image in, text out, via an integrated vision encoder | Yes, natively | Yes, natively |
| Weights location | huggingface.co/meta-models/Muse-Glimmer-30B | Meta developer estate and partners | Meta developer estate and partners |
| Deployment positioning | A single consumer GPU or a Mac; over 55 GB at full precision, under 20 GB quantized | Single H100 GPU with Int4 quantization | Single H100 host, meaning one multi-GPU server rather than one GPU |
| Released | August 10, 2026 | April 5, 2025 | April 5, 2025 |
| 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. Read the Usage Policy Meta attaches on the model page, though, and don’t generalize the permissiveness. Llama 4 hasn’t been relicensed.
· On October 2, 2026 Meta updated its Superintelligence Scaling Framework so it now has to consider how released open weights could be modified. That’s a governance input for future open releases, not a change to Glimmer’s license.
- Two licenses, two risk profiles
- Check which model your vendor actually ships
- Record the license with the artifact
Where the Glimmer weights actually are
Meta names a weights repository on HuggingFace at meta-models/Muse-Glimmer-30B, with companion repositories for GGUF, ExecuTorch and an assistant build. It lists runtime packaging for Ollama, llama.cpp, ExecuTorch, LM Studio and Unsloth Desktop, and names Fireworks, Together.AI and OpenRouter as hosted options. That’s enough to specify a deployment and enough to test one this week.
· Meta’s docs cover serving it with vLLM or SGLang behind OpenAI-compatible endpoints, which makes an A/B against a hosted Muse Spark endpoint easy to set up.
- One main repository, three companions
- vLLM and SGLang paths documented by Meta
- Pin the repository revision, not just the name
Long-context analysis
Scout’s supported 10M context can carry codebase, archive and multi-document workloads that are otherwise split across requests. Muse Glimmer’s 128K default is a different class of job: an always-on agent holding a working session, not a model swallowing an archive. Meta says longer contexts are supported without saying how much longer, so test the ceiling yourself.
- Test retrieval at real lengths
- Build coverage checks
- 128K and 10M are not competing for the same workload
Multimodal understanding
Scout and Maverick accept text and image inputs for grounding, visual question answering and application-specific workflows. Muse Glimmer does too: Meta documents a dense multimodal transformer with an integrated vision encoder, taking text and image inputs and returning text only. So a vision task isn’t a reason to rule Glimmer out.
- Use exact visual tasks
- Glimmer returns text only, no image output
- 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. For Muse Glimmer, Meta points to a Usage Policy rather than a named safety model, so the classification layer is yours to choose.
- One safety layer, not the whole system
- Tune policy to the use case
- Monitor false positives and misses
Repoint every Llama link you hold
Meta’s developer estate sits on dev.meta.ai, and the old llama.developer.meta.com host redirects there. The Llama 4 license lives at dev.meta.ai/llama/llama4/license, with Section 2 in full. The hosted Llama API deprecation route, dev.meta.ai/docs/llama-api-deprecation, has no confirmed content. A redirect isn’t a guarantee the specific document survived.
- Model pages, docs, prices and the license: dev.meta.ai
- Model announcements: research.meta.ai
- Developer programs: developers.meta.com/blog
- Repoint bookmarks, citations and contract annexes
Open models beyond the chat families
Meta’s open-model story is wider than Llama and Glimmer. On March 27, 2026 Meta announced SAM 3.1 as a drop-in replacement for SAM 3. 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. SAM 3.1 is the version in Meta’s model list, under the ID sam-3.1, and Meta states no license for it.
- Vision models, not chat models
- Fine-tuned across 300 A100 GPUs
- SAM 3.1 is the current version, license unstated
The open-weight advantage appears only if the organization can operate it
A downloadable model buys you control, customization and portability. It also hands you a permanent job: serving, security and evaluation, forever. You wanted the weights, right? Then you own what comes with them. Two Meta open families are in play under different licenses, so the artifact you pin has to record which one you deployed. Muse Glimmer has every field a change record needs: an August 10, 2026 release, a weights repository, a context window, an architecture and named runtimes. And the hosted comparison can be priced now, because Meta prints its Model API rates. So the business case is arithmetic, not guesswork.
What you can pin on Muse Glimmer
A pinned artifact needs an identifier, a weights location, a context configuration and a release date. All four are published.
· Identifier: the HuggingFace repository meta-models/Muse-Glimmer-30B. The name muse-glimmer is the Ollama tag, not a Model API ID, because Glimmer isn’t served on the Model API. Weights: HuggingFace, with packaging for Ollama, llama.cpp, ExecuTorch, LM Studio and Unsloth Desktop. Context: 128K by default, longer supported. License: Apache 2.0 plus Meta’s Usage Policy. Released: August 10, 2026.
· Record the repository revision hash beside the date, because the repositories kept updating after release.
- Repository published
- Weights location published
- Context window published
- Release date published
Where the hosted option sits
Meta’s own endpoint serves Muse Wonder and Muse Spark, not Llama and not Glimmer, and Meta states that outright. So a hosted-versus-self-hosted comparison inside Meta’s estate is a comparison between different models, not different deployments of the same one. The cost side is published: $1.25 in and $4.25 out for muse-spark-1.3, $4 and $20 for muse-wonder-1.0. The status side isn’t settled. The recap says generally available. The live product page says public preview.
- Hosted means Muse Wonder or Muse Spark
- Status contested between Meta’s own pages
- Model cost on the standard ID, not the contributor one
The hosted track is multi-cloud
Muse Spark 1.3 is available through Oracle Cloud AI Platform, and Google Cloud has it in private preview. Meta names Fireworks, Together.AI and OpenRouter for Muse Glimmer, and fal, Runway and OpenRouter for Muse Image. So “Meta runs no inference service for open weights” is still true, and “you can only get Muse Spark from Meta” isn’t. That matters for anyone with a cloud commitment they’d rather spend than add a new vendor.
- Oracle Cloud AI Platform: available
- Google Cloud: private preview
- Check data terms per cloud, they are not Meta’s terms
Cloud and edge partners
Llama is available through model hosts, clouds and edge ecosystems. Muse Glimmer’s hosted partners are named by Meta: Fireworks, Together.AI and OpenRouter. Ask any other vendor claiming to serve it exactly which weights they hold, which revision, and under which license.
- Avoid provider-specific lock-in
- Check quantization and context differences
- Document the deployed artifact
Twelve habits for keeping Meta’s tracks, names and IDs straight
Name the product, the model, the ID, the license and the access path. Precisely. That’s still the habit that matters most, and Meta keeps making it harder. It reuses the word Muse for an agent, a coding tool, a small-business product and a device. It publishes open weights under two licenses. It sells the same Muse Spark model under two IDs with two very different data deals. And it has an enterprise division you can’t price. So precision isn’t pedantry here. It’s the whole job.
Say which Muse you mean
On its own the word Muse covers the model family, the consumer agent app, Muse Code, Muse for Small Business, Muse Charm and the Muse API. Name the thing. Muse Wonder is the top hosted model. Muse Spark 1.3 is the cheaper one. Muse Glimmer is the open model. Muse is the consumer agent. Meta AI is the in-app assistant, which isn’t branded Muse. Never write only “Muse.”
Name the license, not just “open”
Muse Glimmer is Apache 2.0 with a Meta Usage Policy. Llama 4 is the Llama Community License, with a user-count condition. Open weight isn’t one thing.
Delete the word contributor
Search every config for -contributor. That suffix trades a 92% input discount for Meta’s right to train future models on your content. If someone chose it on purpose, get that choice signed off. If nobody chose it, remove it now.
Use Muse for distribution
Choose Meta’s hosted track when social, messaging, camera, voice or creator context is the advantage. The price is published. The status isn’t settled, so get it in writing before you call it production.
Use open weights for control
Choose Glimmer or Llama when hosting, customization or data location is the requirement. And budget for the hosting Meta doesn’t provide.
Watch five hosts
Point monitoring at about.fb.com/news, research.meta.ai/blog, dev.meta.ai, developers.meta.com/blog and www.meta.com/blog. And ai.meta.com/blog hasn’t posted since July 27, 2026, so treat quiet there as a channel fact, not a product fact.
Price the subscription
Meta One runs from $2.99 to $499 a month across nine tiers. Muse Code runs from $5 to $200. Meta AI is still free for everyday use, so state which side of that line the client’s workflow falls on.
Match the context window to the job
Muse Wonder and Muse Spark list 1M tokens. Muse Glimmer is 128K by default, with longer supported but no stated ceiling. Scout is 10M. Those are different jobs, not a ranking. Test at the real workload size.
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. And the Muse agent has connectors to payment, commerce and code systems, so the pause matters more.
Operate the open model
Pin, secure, evaluate and maintain every deployed artifact, and record which license it came under. For Muse Glimmer, record the repository revision too, because the repositories kept updating after the August 10 release.
Do not budget against the enterprise platform
Meta Enterprise Platform has a leader, a bundle and a strategic claim. It has no price, no tier and no date. Put it in the roadmap conversation, never in the cost model, until Meta publishes commercial terms.
Reach, personal context, training rights and two open licenses create different risk profiles
The costliest Meta mistake right now is one word in a model ID. Pick the contributor version of Muse Spark and Meta can train on what you send it. Past that, Meta’s risks don’t come in one shape. The consumer apps raise privacy, representation and publishing questions. The open weights hand safety, security and lifecycle work to whoever deploys them, under two sets of terms. The Muse agent reaches payment, commerce and code systems. And Meta’s own pages disagree on whether the Model API is generally available. So you’ll often decide with half the vendor facts. Name exactly which half is missing.
Contributor IDs hand Meta training rights
Meta labels muse-spark-1.3-contributor and muse-spark-1.2-contributor “Used to improve our products.” Its pricing docs say contributor mode gives discounted prices in exchange for permission to use your content to train future Meta models. The plain IDs, muse-spark-1.3 and muse-spark-1.2, are labeled “Not used to improve our products.” The contributor input price is $0.10 per million tokens against $1.25. Cheap enough that someone will pick it to hit a budget. So make the ID a reviewed setting, not a developer default.
Model API status: Meta’s pages disagree
The September 24, 2026 Connect recap says generally available globally. The live product page says public preview. The Muse Spark 1.1 page says public preview for US developers. The recap records what Meta announced. The live page is the better guide to today, and it keeps no history, so save a dated copy. So don’t contract on “generally available” until Meta states the status and its support terms in writing for your account.
Meta Enterprise Platform has no commercial detail
This is the biggest gap in Meta’s AI estate. And it’s Meta’s omission, not a documentation problem. The September 28, 2026 announcement names a division, a leader and four bundled components. That’s it. No pricing, no tiers, no availability dates, no API specifications. Zuckerberg calls it the next major pillar of the business. That raises the stakes. It doesn’t fill the gap. So if your procurement process needs a Meta enterprise answer this quarter, give it the Model API, Muse Code plans and Meta One, and note this one as pending.
Llama license: the 700 million user clause
Open weight isn’t the same as unrestricted open source. And this is the clause that decides it. Here it is in plain words. Did your products have over 700 million monthly active users in the month before the Llama 4 release date? Then you must ask Meta for a license. Meta can refuse. And until it says yes, you get no rights under the agreement at all. Muse Glimmer, under Apache 2.0, carries no equivalent condition.
· Source: Section 2 of the Llama 4 Community License Agreement, effective April 5, 2025, at dev.meta.ai/llama/llama4/license.
· The trigger, verbatim: “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.”
· The consequence, verbatim: a license “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.”
Apache 2.0 is the exception, not a policy change
One Meta open model is 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, and attached a Usage Policy to that one too.
Muse Glimmer is not a Model API model
It’s easy to assume an ID like muse-glimmer means Meta hosts it. Meta doesn’t. Its docs say Glimmer runs on your own hardware and isn’t called over the Model API, and the models list endpoint leaves it out. That name is only the Ollama tag. So there’s no Meta token price, no Meta rate limit and no Meta uptime behind it. Your serving stack, or a partner’s, carries all of that.
A Muse Spark 1.1 test reached the real internet
On August 14, 2026 Meta disclosed that while Irregular was testing a pre-release Muse Spark 1.1 with safeguards removed, a misconfiguration let the model reach the open internet, where it found and exploited a vulnerability in a real website and changed its database. Meta reviewed more than 10,000 activity records, found no other case, called it neither a sophisticated attack nor a sandbox escape, and now bars test scenarios that reference real companies.
· Meta promised a full safety report for Muse Wonder that isn’t out yet. Ask for it before high-autonomy agent work.
Muse connectors are a new data surface
The Muse agent connects to Walmart, Best Buy, American Eagle Outfitters, DICK’S Sporting Goods, Fanatics, Gap, Michael Kors, Sephora, Ulta, Wayfair, Shop Pay, PayPal, Expedia, Instacart, Notion, Granola, GitHub and Box. Meta also says Muse gets its own email address it can use to get things done. Payment rails, a code host, a document store and a mailbox, all reachable by an agent. Muse for Small Business adds Shopify, Stripe, Slack, Figma, Canva, Asana, Dropbox, Zoom and Intuit QuickBooks among others, with a stated approval gate before anything publishes, sends or spends. Review the approval model before anyone connects a work account.
The hosted Llama API shutdown date is still unsourced
The July 6, 2026 shutdown date has no first-party citation. Meta’s deprecation route at dev.meta.ai/docs/llama-api-deprecation confirms nothing, and dev.meta.ai/llama carries no date and no shutdown notice. What Meta does state is the absence: Llama models aren’t run for you through the Model API. That makes the shutdown far more likely and still doesn’t date it. So if an architecture decision depends on the date, check it before the decision, not after.
The developer estate keeps moving
The old llama.developer.meta.com host redirects to dev.meta.ai. Developer programs post on developers.meta.com/blog, note the s, and model announcements moved to research.meta.ai. Those are new hosts, not new paths on old hosts. So repoint every Meta reference in your procurement pack, contract annexes and internal wiki. And don’t read a redirect as proof the document survived.
Muse geography: two Meta pages, two answers
The September 8, 2026 launch article, revised September 30, says Muse is rolling out in the US on iOS, Android and muse.ai. The Muse for Small Business article, revised October 8, 2026, calls Muse available in the US and Canada. So don’t promise Canadian staff access on the strength of either page alone. Test it from a Canadian account and keep a dated copy of both articles.
Meta One cost and limits
Nine tiers from $2.99 to $499 a month, global since September 15, 2026. Meta states Meta AI stays free for everyday use. But its announcement defines neither everyday use nor per-tier AI limits. So a client planning volume media generation should establish the ceiling inside the product before buying, or cost it on the Model API instead. Hearing Enhancement rides on the plan too, which pulls a regulated capability into a consumer subscription decision.
Hardware dates are announcements, not shipments
Meta Adventurer goes on sale October 23, 2026. Ray-Ban Meta Audio was set to ship October 13, and the only Meta page saying it’s on sale carries no date. Meta VR Glasses is $1,299.99 in spring 2027, a season, not a date. Hearing Enhancement is “later in 2026” with no date. Muse Charm has a product page with no announced price or date. Muse on glasses is “coming months.” Sort any hardware dependency into those buckets before it reaches a budget.
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. Forum, the Facebook Groups app Meta launched on September 29, 2026, adds an AI Ask feature that surfaces existing group posts and comments, which is the same signal with a new front door.
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. Meta says Muse is for users aged 18 and over, and the Glimmer model card says it isn’t intended for users under 18. Check the specific market and the specific product, not the headline.
Intention is not capability
The August 10 letter promised to resume open source releases, and that promise has been kept. Its other two promises are an exceptionally capable personal agent for everyone, and a fully private mode where even Meta can’t see or grant access to your information. Meta hasn’t announced either as shipped. 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.
Direct sharing
The shortest path from generation to publication increases accidental or insufficiently reviewed output.
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. Video generation ships inside Meta One, but Meta attributes it to Muse models as a family and publishes no video model ID, so don’t call it Muse Video. Muse Realtime Avatar is a named research system, not a product you can buy. And Meta Enterprise Platform is a division, not a purchasable product.
Where these facts come from
Three groups do the work. First, dated posts on research.meta.ai and dev.meta.ai’s developer blog. They date Muse Wonder, Muse Spark 1.2 and 1.3, Muse Glimmer, Muse Code and Muse Voice Transcribe. Second, dated newsroom posts on about.fb.com for the enterprise division, Muse for Small Business, Muse Wonder and the hardware. Third, undated pages on dev.meta.ai. They’re the only record of Model API prices, contributor terms, model IDs and the public preview label, so each one is stamped with the day it was read. And two things still have no first-party answer: the July 6, 2026 hosted Llama API shutdown date and Meta Enterprise Platform pricing.
AI Mindset