News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgfnMuse

Muse

Meta's multimodal LLM family, successor to Llama, with open and closed variants

Muse is a large language model family from Meta, built by Meta Superintelligence Labs since April 2026 as the successor to Llama. It spans closed models (Spark, Wonder, Code) and open weights (Glimmer, Spark Open), with multimodal reasoning, coding and long-context use.

Share on WhatsApp

  • Successor to Meta's Llama family, in development since April 2026
  • Spark Open: 816B total parameters, 42B active, mixture of experts
  • Glimmer: 30B open-weight multimodal model, runs offline on a 24 GB GPU
  • Spark handles up to one million tokens of context
  • Spark 1.2 served via Meta Model API and OpenRouter

What it is

Muse is a family of large language models made by Meta through its Meta Superintelligence Labs unit, active since April 2026. It replaces the Llama line that Meta published between 2023 and 2025. The family mixes closed models, such as Spark, Wonder and Code, with downloadable open weights, Glimmer and Spark Open. Spark is built for multimodal reasoning, coding and AI-assisted software work, and also backs other Meta AI products. It can read up to a million tokens in one go. Spark Open, the released checkpoint, is a mixture of experts design with 816 billion parameters in total and 42 billion used per token.

How it came about

Meta Superintelligence Labs began work on Muse in April 2026 and shipped four Spark releases that year: 1.0 on April 8, 1.1 on July 9, 1.2 on August 5 and 1.3 on September 2. August also brought Muse Code and cheaper Contributor tiers tied to giving up privacy rights over submitted data. Muse Wonder arrived on October 12 as a rival to the largest models from Anthropic and OpenAI. Open weights followed: the small Glimmer model on August 10 and the full-size Spark Open checkpoint on October 15.

Inside the model

The open Spark Open checkpoint is a mixture of experts model, meaning only part of the network works on each token: 816 billion parameters overall, 42 billion active. Glimmer is a 30-billion-parameter causal language model that pairs a large language model with a separate perception encoder, so it accepts both words and images. It targets local agentic and coding jobs and runs offline on a 24 GB consumer GPU. Spark itself handles a million tokens of context. Contributor tiers let Meta keep submitted data to improve its products; the pricier standard tier does not.

What it can do

Muse is aimed at multimodal reasoning, coding and AI-assisted software development, and also serves other Meta AI products. Muse Code, a terminal-based coding agent harness released with Spark 1.2, writes and debugs code, handles longer development tasks and runs several sub-agents at once. Wonder Full, a stronger version of Wonder, sits behind the top subscription tier and is not offered through the API. No benchmark figures are given here.

How to use it

Spark 1.2 is served through the Meta Model API and OpenRouter. Muse Code is a terminal tool. Contributor tiers cost less but let Meta retain submitted data; the standard tier does not retain it for product improvement. Glimmer's weights ship under Apache 2.0, and Spark Open uses the Muse Model License. Spark and Wonder are proprietary. Wonder Full needs the top subscription and has no API.

Safety and testing

Contributor tiers are tied to users waiving privacy rights over the data they submit. A September 2026 report described Meta testing Muse agent calls that were in fact handled by people in a call centre.

Reception and debate

Wonder was framed at launch as a competitor to Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra. Reporting in September 2026 covered Meta trialling human call-centre staff behind some Muse agent calls.

Why it matters

Muse succeeds Llama, the open weights family that Meta released from 2023 to 2025, and carries that line forward with both closed and downloadable models. It also underpins other Meta AI products and services.

Questions people ask about Muse

What is Muse?

Muse is Meta's family of large language models, built by Meta Superintelligence Labs from April 2026. It replaces Llama and mixes proprietary models like Spark and Wonder with openly released ones such as Glimmer and Spark Open.

Which Muse models can I download?

Two: Muse Glimmer, a 30-billion-parameter multimodal small model under Apache 2.0 released in August 2026, and Muse Spark Open, an 816-billion-parameter mixture of experts checkpoint released in October 2026 under the Muse Model License.

How big is the largest open Muse model?

Muse Spark Open has 816 billion parameters in total, of which 42 billion are active for each token because it is a mixture of experts design.

How do I use Muse Spark?

Spark 1.2 is available through the Meta Model API and OpenRouter, and Muse Code is a terminal-based coding agent that runs on the model. Contributor tiers are cheaper but let Meta keep your submitted data.

What can Muse Code do?

It is a terminal harness released alongside Spark 1.2 in August 2026. It writes and debugs code, works through longer software development tasks, and can run several sub-agents at the same time.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Muse (language model)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.