1. Home
2. Guides
3. Meta

Meta Muse models explained: Muse Spark 1.3, 1.2, 1.1 and Glimmer, with release dates

Meta's frontier models stopped being called Llama in April, and the only Muse you can download so far is the small one.

By Think Facility · September 23, 2026 · Meta Meta AI models

The newest Meta model is Muse Spark 1.3, released on September 2, 2026. It's the fourth Muse Spark since April, and Meta's roadmap for it includes something the earlier ones never had: an open-weights release, not yet dated.

If you last looked at Meta's AI when it meant Llama, most of this is new. Meta's frontier work now comes out of Meta Superintelligence Labs under the Muse name, and the newest repositories in its Llama account on Hugging Face date from April 2025. I pieced the lineup below together from Meta's own launch posts, its developer pages and the Muse Glimmer model card, all read on September 23, 2026.

Every Muse release, by date

Dates from Meta's launch posts on ai.meta.com, research.meta.ai, about.fb.com and its developer blog, read September 23, 2026.

Where Llama went

Muse Spark launched on April 8, 2026, and Meta was clear it was a break from what came before. Meta Superintelligence Labs says it spent nine months rebuilding its pre-training stack, and that the new recipe reaches the same capability with "over an order of magnitude less compute than our previous model, Llama 4 Maverick". Pre-training is the expensive first stage of building a model.

That first Muse Spark was built for Meta's own products. It went into the Meta AI app and website, with API access limited to a private preview for selected partners. Developers couldn't sign up until July.

Meta's developer site still carries Llama 4 docs, but every model it has launched since April carries the Muse name.

Muse Spark 1.1, 1.2 and 1.3

The version numbers here move fast, and the three are priced identically, which makes them easy to confuse.

Muse Spark 1.1, on July 9, was the one that opened the door to developers. Meta launched its Model API in public preview alongside it, with a million-token context window. Meta pitched it at agent work: planning a job, handing parts of it to subagents and operating a computer.

Muse Spark 1.2, on August 5, arrived with Muse Code, Meta's first coding agent, in early beta. Meta called 1.2 "a moderate improvement over Muse Spark 1.1", tuned for the jobs coding agents get most, like multi-file refactors and long debugging sessions. The same day, Meta widened its API preview to developers worldwide.

Muse Spark 1.3, on September 2, is the current one. Meta says it was trained on more long-running coding tasks, asks clarifying questions when a prompt is ambiguous and confirms before taking consequential actions. In comparisons by Meta's own engineers it used about 20% fewer tool calls and 25% fewer tokens than 1.2 on the same work. Fewer tokens at the same price is a real saving, if it holds up on your workload.

All three are still listed on the Model API at $1.25 per million input tokens, $0.15 for cached input and $4.25 for output. Meta still lists 1.1 and 1.2 beside 1.3, so there's no forced move yet.

How you call Muse Spark

There are three ways in. Meta's Model API is the direct one, and its page says you can point code written for the OpenAI SDK at it with a new key and address, which spares you a rewrite. Muse Spark is also on OpenRouter, the service that resells many labs' models behind one account. And Muse Code, Meta's terminal coding agent, installs with one command and runs on the same API.

The API can also search the web for you. Meta offers search grounding as a tool you add to a request (grounding meaning the answer is built from pages it fetched just then), which matters for anything that changed after the model was trained.

The contributor tier

This was the detail I didn't expect on a price list. Muse Spark 1.2 and 1.3 each come in a second version with -contributor on the end of the ID, priced at $0.10 in and $0.20 out, under a twentieth of the standard output rate. Meta's table marks those IDs "Used to improve our products" and the standard ones "Not used to improve our products".

So the cheap tier is cheap because your prompts help train Meta's models. That's a fair trade for a hobby project and a bad one for client code. Muse Code starts new users on the contributor tier, rate-limited by tokens over a rolling five-hour window, and Meta's 1.2 post says to swap the model ID to the standard one when you need more. We've gone through Muse Spark 1.3's limits in Cursor in another piece.

Muse Glimmer, the one you can download

Muse Glimmer came out on August 10, 2026 with its weights on Hugging Face under Apache 2.0, a license with no user or revenue conditions. It's a 30-billion-parameter model (29.6 billion, counting a 1.8-billion-parameter image encoder) that Meta says was distilled from Muse Spark, meaning trained to imitate the bigger model's answers.

Meta built it for agents that run entirely on your own machine. Its model card says a roughly 4-bit version of the language model fits under 20 GB, leaving room for the rest inside a 24 GB or 32 GB graphics card, and the full-precision version wants 64 GB. It reads 131,072 tokens and its knowledge stops at January 4, 2026.

It also ships with a small companion model for speculative decoding. The card says this drafter proposes 16 tokens at a time, and Meta measured the pair on M4 Max and M5 Max MacBooks and on an Nvidia RTX 5090. Ready-made GGUF and ExecuTorch versions sit beside the main repository, so llama.cpp users don't have to convert anything.

Note where it lives. The Hugging Face organization is meta-models, a new account, so searching Meta's old Llama account turns up nothing.

Muse Image, Muse Video and the Muse agent

Muse Image launched on July 7 inside Meta AI, Instagram and WhatsApp, and it's on the Model API at $0.01 an image. Muse Video was previewed in the same announcement and is still described as "coming soon to creators and Meta AI".

Then there's Muse with no second word. Launched on September 8, 2026, it's a personal agent app that runs on a dedicated virtual machine Meta calls Muse Secure VM, works in its own app or inside WhatsApp, and is powered by Muse Spark. We covered its arrival on the Mac separately.

Muse Spark 1.3's announcement ends with Meta's roadmap: "bigger models, the Muse Spark open weights release, and more."