Meta AI: Llama, Muse Models and Weight Access

Meta develops Llama and Muse. Glimmer has Apache 2.0 weights; Muse Spark Open has downloadable weights under Meta’s separate Muse Model License.

kindCompany
foundedJul 2004

Latest

Updated

Models and the products around them

Meta Platforms, Inc., formerly Facebook, Inc., operates consumer services and develops AI models. Its filing records incorporation in July 2004 and the corporate name change in October 2021. The Llama releases gave developers weights they could run and adapt under Meta's applicable terms, including commercial access with Llama 2 in 2023.

Meta's June 2025 Scale AI investment also brought Scale founder Alexandr Wang to Meta's AI effort. Scale's announcement explicitly described a minority equity position and continued independent operation. That relationship supplied talent and data collaboration, rather than establishing that Meta bought the entire Scale company.

Different releases have different access terms

Muse Spark arrived in April 2026 as the first model from Meta Superintelligence Labs, initially through hosted access. Muse Glimmer followed in August with downloadable 30-billion-parameter weights under Apache 2.0, using distillation from Spark for a smaller model aimed at local agent workflows.

On October 15, Meta released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters. The Spark family now included a downloadable variant alongside its hosted offerings. The release also listed hosted API access separately from weight downloads.

Spark Open uses the Muse Model License, rather than Glimmer's Apache license. Section 3 requires a separate agreement for a model-as-a-service business when the licensee and its affiliates exceed $20 million in aggregate gross revenue in any consecutive twelve months. It expressly excludes internal use and commercial activity that is not model-as-a-service. Access and licensing therefore need to be read per release.

Sources

  1. [1]
    Meta Platforms, Inc. Form 10-K for fiscal 2021

    Meta Platforms, Inc. · Feb 3, 2022

  2. [2]
  3. [3]
    Muse-Spark-Open-816B-A42B

    Meta AI Research · retrieved Oct 23, 2026

  4. [4]
  5. [5]
  6. [6]
  7. [7]

Meta history

2026

│
Meta releases Muse Ember, its smallest Muse model

Muse Ember 1.0 is a model for high-volume agent work such as tool calls, routing and extraction, priced at $0.10 and $0.50 per million input and output tokens with a 1 million token context window. It went live on the Meta Model API and in Muse Code.

Meta
Model
│
Meta pauses Muse conversation sharing after a privacy report

With Claroc

Meta suspended Muse's sharing feature after Claroc found more than 6,800 share hostnames in passive DNS records. Researchers opened a sample without logging in and found some personal and internal business documents; Meta said it was reviewing how the links could be discovered.

Claroc
Product
│
Meta shareholders approve non-voting Class C stock

Meta's controlling shareholders approved a new non-voting share class, and Mark Zuckerberg signed a founder agreement, on October 20. Disclosed on October 21 as a financing tool for the company's AI expansion, the recapitalization schedules a one-for-one stock dividend after the charter takes effect in January 2027, preserving existing shareholders' voting power.

Meta
Company
│
Meta releases Muse Spark Open with downloadable weights

An 816 billion parameter mixture-of-experts model, 42 billion active, under Meta's own license: a licensee whose revenue with affiliates passes $20 million in any 12 months must notify Meta and sign a separate agreement before offering it as a hosted service. Meta billed it as the strongest US open-weight model; its own card has it trailing Kimi K3 on the Artificial Analysis index, 39 to 44.

Meta AI Research
Model
■
Meta releases Muse Wonder and Muse Wonder Full

Meta's superintelligence lab followed Muse Spark with larger models built for agents that carry memory across conversations and keep working through blocked tasks. Muse Wonder went on Meta's Model API at $4 input and $20 output per million tokens; the larger Wonder Full stayed inside Muse Code subscriptions.

Meta AI Research
Model
│
Biohub's Virtual Biology Initiative grows to about $1.8 billion with US and AI lab money

With Biohub, United States Department of Energy, Google DeepMind, Isomorphic Labs

Biohub's Virtual Biology Initiative grew to about $1.8 billion in funding, data, computing and instruments, including more than $500 million over five years from the Energy Department and $300 million from Google DeepMind, Isomorphic Labs and Meta. Biohub's own April pledge and existing NIH data make up much of the rest.

Biohub
Deal
│
Meta adds small business tools to Muse

Meta added business skills and connectors to its existing Muse agent, including access to Facebook and Instagram business accounts and tools such as QuickBooks, Shopify and Stripe. Users can connect available services now; Meta says publishing, sending and spending still require approval.

Meta
Product
■
Meta launches Muse, a personal AI agent

Muse books appointments, shops, and works through multi-step tasks from a dedicated cloud VM with its own browser, built under chief AI officer Alexandr Wang. Meta shipped it with a free tier plus $20 and $100 monthly subscriptions on iOS, Android, and the web, and said an end-to-end encrypted VM option is coming later this year.

Meta
Product
│
Meta ships a terminal coding agent

Muse Code entered beta on Muse Spark 1.2, with parallel subagents, worktree isolation, and a crash-safe event log. The first coding product out of Meta Superintelligence Labs, priced under the incumbents at $1.25 and $4.25 per million tokens.

TechCrunch
Product
■
Lab employees ask Washington for a brake they can pull later

With OpenAI, Anthropic, Google DeepMind

More than a thousand people inside the frontier labs, including Anthropic's chief executive and OpenAI's chief scientist, asked the United States to back an international effort to build the tools for deliberately pacing automated AI development. The statement does not ask anyone to slow down now.

Pacing the Frontier
Culture
│
Google limits Meta's access to its Gemini models

With Google

Meta had been buying Gemini capacity for safety filtering, coding assistants and internal tools; Google could not fill the order and capped it, delaying some of Meta's own AI projects. Even a company building its own frontier models still needed someone else's compute.

CNBC
Deal
■
China orders Meta to unwind its purchase of Manus

With Manus, National Development and Reform Commission

The National Development and Reform Commission prohibited the $2 billion acquisition of the Singapore-based, Chinese-founded agent startup and told the parties to withdraw. Beijing had started policing outbound AI talent the way Washington polices chips.

CNBC
Policy
■
Meta returns to the frontier with Muse Spark

The first model out of Meta Superintelligence Labs after nine months of rebuilding the stack, multimodal, with instant and thinking modes and subagents that run in parallel. It shipped closed, which for the company that published Llama was the news.

Meta
Model
│
SAM 3.1 tracks sixteen things at once

Object multiplexing puts up to sixteen tracked objects through a single forward pass, taking a mid-density video from 16 to 32 frames per second on one H100. Checkpoints went to Hugging Face.

Meta
Model
│
Arm launches its own AGI data-center CPU with Meta as co-developer

With Arm, Super Micro Computer

Arm introduced its AGI CPU, expanding from licensed processor designs into Arm-designed data-center silicon. Meta was the lead co-developer and Supermicro contributed a liquid-cooled system design; early systems were available, with a broader rollout expected later.

Arm
Hardware

2025

■
Meta buys Manus for more than $2 billion

With Manus

The Singapore agent company had gone from launch to more than $100 million of annual recurring revenue in eight months. Meta's condition was no continuing Chinese ownership and a wind-down of Manus operations in China, which is the thread Beijing pulled on four months later.

Euronews
Acquisition
│
Segment Anything learns to take a description

SAM 3 finds and tracks every instance matching a phrase like red baseball cap, where SAM 1 and SAM 2 returned one object per prompt. Meta released the weights, an open-vocabulary benchmark and a paper, along with SAM 3D for rebuilding objects and human bodies from a single image.

Meta
Model
│
Yann LeCun leaves Meta to bet against language models

Meta's chief AI scientist departed after twelve years to start a company on world models, having spent much of that time arguing publicly that scaling language models would not reach human-level intelligence. Meta had just spent $14.3 billion building a lab around the opposite thesis.

CNBC
Company
│
Meta cuts 600 jobs from its AI Superintelligence Labs

The cuts came months after Meta paid large packages to recruit more than 50 researchers into the unit. Chief AI officer Alexandr Wang said a smaller team would need fewer conversations to reach a decision; most affected employees were expected to find other roles inside the company.

TechCrunch
Company
│
Meta launches Vibes, a feed of AI-generated video

The feed sits inside the Meta AI app and meta.ai, built on image models from Midjourney and Black Forest Labs, and lets people generate, remix and cross-post short AI videos to Instagram and Facebook. Commenters mocked it as a river of AI slop.

Meta
Product
■
Meta buys half of Scale AI and a chief AI officer

With Scale AI

A $14.3 billion investment for 49 percent of the data-labeling company, whose founder Alexandr Wang left to run a new superintelligence lab inside Meta. Talent had become the thing you buy a company to get.

Scale AI
Acquisition
│
Meta releases V-JEPA 2, a world model trained on video

V-JEPA 2 learns physical intuition from video to help robots predict how the world responds to their actions, reaching state-of-the-art results on visual understanding and prediction benchmarks. Meta released the weights under a non-commercial license alongside new benchmarks for physical reasoning.

Meta AI
Research
│
Meta holds its first LlamaCon developer conference

Meta's first event built around its open Llama models, held in Menlo Park. Announcements included a Llama API in limited preview, new safety tooling and inference partnerships with Cerebras and Groq.

Meta
Culture
│
Meta releases Llama 4 Scout and Maverick

Two mixture of experts models with open weights, published on a Saturday to a reception Meta had not planned for. It was the last Llama release before the company rebuilt its AI organization around a new lab.

Meta
Model
■
Zuckerberg commits Meta to $60-65 billion in AI capex for 2025

Mark Zuckerberg said Meta would spend $60 billion to $65 billion on capital expenditures in 2025, mostly data centers and chips for AI, a roughly two-thirds increase over 2024. He said the buildout was meant to make Meta AI the most-used assistant of the year.

Fortune
Deal

2024

│
Meta previews Movie Gen, a video and audio generation model

A 30 billion parameter model turns text into 16 second videos, personalizes them from a photo, edits footage with plain language commands and generates matching audio. Meta called it research rather than a product and said it would work with filmmakers before any release.

Meta AI
Research
│
Frontier labs sign the Seoul safety commitments

With OpenAI, Anthropic, Google, Microsoft, Amazon, Mistral AI, xAI, Cohere, G42, IBM, Inflection AI, Naver, Samsung Electronics, Technology Innovation Institute, Zhipu AI

At the UK and South Korea's Seoul summit, sixteen frontier labs including OpenAI, Anthropic, Google, Meta and Zhipu AI committed to publish safety frameworks with risk thresholds and to hold back models that breach them. Voluntary, and the signatory list kept growing afterwards.

Department for Science, Innovation and Technology
Policy

2023

●
Llama 2 ships with a commercial license

With Microsoft

Downloadable models for research and commercial use under a community license, with acceptable-use rules and additional terms for very large services. Access to weights had become a distribution strategy, with conditions attached.

Meta AI
Model

2022

│
PyTorch Foundation forms under the Linux Foundation

With PyTorch Foundation, Linux Foundation

The Linux Foundation announced the PyTorch Foundation as a neutral institutional home for PyTorch, which originated at Meta. The change broadened governance and support around the open-source framework rather than releasing a new model or framework version.

Linux Foundation
Company

2017

│
Microsoft and Facebook announce ONNX model interchange

With ONNX, Microsoft

Microsoft and Facebook announced the Open Neural Network Exchange, a shared model representation intended to improve interoperability across AI frameworks. This initial ONNX announcement preceded the separate version-one release in December.

Microsoft
Open source

2016

│
Technology companies establish Partnership on AI

With Partnership on AI, Amazon, Google, DeepMind, IBM, Microsoft

Amazon, Facebook, Google, DeepMind, IBM and Microsoft established a nonprofit forum for AI research and practices. The Partnership on AI was publicly announced in New York on September 28.

DeepMind
Company