Muse Spark Open

The short answer

Muse Spark Open is the open weight version of Meta’s Muse Spark model, released on October 15, 2026, and the weights are free to download. It is a mixture of experts model with 816 billion parameters in total, of which 42 billion are active on any one token, a context window of 1,048,576 tokens, and native multimodal input across text, images, audio and video. Meta calls it the most capable open weight model to come out of the United States to date. The catch is size. Running it yourself takes about a terabyte of memory, and Meta’s own desktop demo links two Mac Studios that are not on sale yet. For almost everyone, the realistic door is the hosted APIs that went live the same day, from providers including Together AI, Groq and OpenRouter. This guide covers the specs, the three ways in, the license rules that decide commercial use, and the point that makes this release different for readers of this site: the model travels worldwide from day one, even though the Muse app is still limited to the United States and Canada.

What is Muse Spark Open?

Muse Spark Open comes from Meta Superintelligence Labs, and it arrived on October 15, 2026, three days after the closed Wonder models made the headlines. To place it, start with the model it shares a base with. Muse Spark is the closed model behind the Muse app and Meta AI: the personal agent that answers quickly, works across your apps, and runs cheaply enough to offer on a free tier. Spark Open is that same foundation, rebuilt for a very different job. Instead of living inside Meta’s products, it is published as files anyone can download, host and build on, and Meta says it carries extra safety training precisely because it will run without Meta’s own system of guardrails around it.

The documented specifications, gathered from Meta’s announcement and the model card, are these:

• Total parameters: 816 billion, in a mixture of experts design with 256 experts, of which 8 are active per token
• Active parameters: 42 billion per token, so it answers at roughly the speed of a much smaller model
• Context window: 1,048,576 tokens, just over one million
• Inputs: text, image, audio and video, handled by a dedicated perception encoder. Output: text only
• Reasoning effort: adjustable from low to a maximum setting, and set to maximum by default
• Training data: text, images, audio and video, covering more than 100 languages
• Built for: long agentic coding, tool use and multi step reasoning

What Meta chose to show it doing tells you the intended job. In the launch demos, Muse Code powered by Spark Open shipped real changes to Ice Cubes, an open source Mastodon client: bug fixes, new features and a full build. The same setup drove DaVinci Resolve, the professional video editor, reviewing footage to find the strongest moments and cropping it. This is a model aimed at work that stretches across many steps and many files, not at quick questions.

Open weights change who holds the controls. A team can download the model, inspect it, fine tune it on its own material, and host it on infrastructure it chooses, which decides where its data goes. That is the opposite of the Wonder arrangement, where the model stays on Meta’s machines and you rent access by the token. It is also why the license section below matters: downloadable is not the same as unrestricted.

How can you actually use it?

There are three routes in, and they suit very different people. The download is free. Everything after the download is where the routes split.

Route one: download the weights

The weights are published on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta ships two checkpoints. The full precision BF16 build is about 1.6 terabytes of files. The compressed NVFP4 build is about 530 gigabytes. On Hugging Face, the download asks you to sign in and accept the license terms first. The model card names the frameworks ready to serve it, including SGLang, vLLM, TensorRT-LLM, llama.cpp, Unsloth and Apple’s MLX, so the tooling around it is the tooling the open model world already uses.

Route two: run it on your own hardware

This is where the free model gets expensive. TechCrunch, citing Engadget, puts the local requirement at about 1 terabyte of memory. Meta’s model card says a self hosted run on graphics processors wants at least 700 gigabytes of video memory for the NVFP4 build, spread across a rack of top end cards. The desktop version of the same story is the one Meta and Apple told together: the model was optimised for Apple’s MLX framework and runs across two Mac Studios with M5 Ultra chips and 512 gigabytes of memory each, linked so the pair behaves like one machine with a terabyte of unified memory. Two problems follow. The 512 gigabyte Mac Studio is not shipping yet; Apple expects it later in October 2026. And the smaller 256 gigabyte version already costs over $10,000, so the pair in the demo is a workstation budget, not a hobby one.

One softer landing exists, and it is worth labelling honestly. Within days of the release, community builds from groups such as Unsloth had compressed the model toward 2 bit precision, and community reports describe those builds running on a single machine with about 256 gigabytes of memory through llama.cpp. Treat that as an experiment reported by the community, not a setup Meta supports: compression at that level costs quality, and nobody has validated it at scale yet.

Route three: call a hosted API

For almost everyone, this is the route. Seven providers had Spark Open live on launch day: Together AI, Baseten, Fireworks AI, Groq, Modal, Nebius and OpenRouter. You sign up, take an API key, and pay per token, with no download and no hardware. Pricing varies by provider, and each provider sets its own rates, so check the provider’s own page before budgeting rather than trusting any roundup, including this one. One cost lever is built into the model itself: its reasoning effort setting decides how long it thinks before answering, it sits at maximum by default, and you cannot switch the thinking off entirely. Setting it to low for simple jobs is the difference between a sensible bill and a silly one. New to paying by the token? Our Muse tokens explained guide walks through how token billing works.

What does the license allow?

Spark Open is not open source, and it does not use a standard license like the Apache 2.0 terms on the smaller Glimmer. It ships under a custom Muse Model License. As reported from the license text by VentureBeat and others, the license grants the rights developers actually need: you may run the model, modify it, fine tune it, distill it and redistribute it, subject to its conditions and a usage policy. Meta states that it claims no ownership of the outputs the model generates for you.

Two obligations decide most real world cases. The first is attribution. Products built on the model must carry the words “Built with Muse Spark Open” in a permitted place, which the license allows to be the documentation, a website or the settings. The second is the revenue line, and it is widely misunderstood, so here it is carefully. If you and your affiliates take in more than $20 million in gross revenue across any twelve month stretch, and you offer Spark Open or a derivative of it as a model service, you must notify Meta and sign a separate written agreement before doing so. The license, as VentureBeat’s reading stresses, explicitly exempts internal deployment, research, commercial activity outside model serving, and use inside your own product when that product is not itself a model service. In plain terms, the $20 million figure is not a ceiling on free commercial use. It is a line aimed at companies reselling the model itself. The license also defines derivative models broadly, broadly enough to cover models improved using Spark Open’s outputs, and it carries usage restrictions, including around military and critical infrastructure use.

The practical summary: almost every developer and business reading this page can use Spark Open commercially, free, today, with an attribution line as the main visible duty. If your plan is to sell access to the model itself, read the license in full before you build.

How is it different from Muse Spark, Wonder and Glimmer?

Four names now carry the Muse brand, and launch coverage keeps smearing them into one. Keep them apart and every claim about this family gets easier to judge.

Muse Spark

The closed model behind the Muse app and Meta AI. It is hosted only, tuned as a fast personal agent, and it is what subscribers and free users alike are talking to in the app today. You cannot download it, and there is no sign that changes.

Muse Wonder and Wonder Full

The closed frontier tier, announced on October 12, 2026. Wonder is reachable through the Meta Model API at $4 per million input tokens and $20 per million output tokens, and inside Muse Code. Wonder Full has no API at all and lives only in the top Muse Code plans. Neither has downloadable weights. Our Muse Wonder guide breaks down both models, their prices and their benchmark record in full.

Muse Spark Open

This page’s subject: the open weight sibling of Spark, released October 15, 2026. Same base family as the app model, extra safety training for life outside Meta’s walls, and access through downloads and third party APIs rather than through any Meta product. It is not in the Muse app, and downloading it does not change what the app runs on.

Muse Glimmer

The smaller open weight model, released on August 10, 2026. Glimmer has 30 billion parameters, was distilled from Muse Spark, and ships under the permissive Apache 2.0 license. It was purpose built for consumer hardware and runs on a single graphics card with 24 gigabytes of memory. That makes Glimmer the honest contrast in this family: if you want a Muse model actually running on your own desk today, Glimmer is the realistic one. Spark Open is the one you rent by the token, or run on a rack.

What does it mean if you live outside the United States?

Start with what did not change. The Muse app launched in the United States in September 2026 and added Canada later that month, and it has not officially launched anywhere else. If you live in India, or in most of the world, you still cannot download the app, and this release does not alter that by one country. Our guide to using Muse outside the US tracks the app situation country by country.

What changed is the model layer underneath. The weights sit on global catalogues, and the hosted providers sell API access across many countries, with each provider setting its own country list and payment rules. So a developer in India can, today, download a Muse model or call one by the token and build a real product on it, without waiting for the app launch and without anyone’s invite code. A business outside the launch countries can evaluate the same model family its future agent will run on, test agentic coding on its own codebase through a hosted API this week, and arrive at the app launch already knowing what the technology does.

Hold the two facts apart, because launch week coverage will blur them. Spark Open being available in your country does not mean Muse has launched in your country. The app, with its invite codes, its waitlists and its consumer plans, is exactly where it was on October 14. What opened is a door for builders, and for the countries still waiting, it is the first Muse door that has opened at all.

What do the benchmarks actually say?

Every headline number attached to this launch is Meta’s, published in Meta’s own comparison tables on October 15, 2026, and no independent benchmark run existed at the time of launch. Read the figures below as claims with a source, not as results this site has verified.

On agentic work, Meta’s table is genuinely strong. MarkTechPost, reporting Meta’s published comparisons, puts Spark Open at 89.2 on MCP Atlas, a benchmark of tool using agents, the top score in that table. The coding picture in the same table is more modest: on Terminal-Bench 2.1, Spark Open scores 76.4, behind Kimi K3 at 85.0, GLM 5.3 at 83.9 and Qwen3.8 at 82.0. Particle’s summary of the launch makes the same point in words: strong agentic and tool use results, with coding scores that trail some leading Chinese open models.

The one independent yardstick available at launch points the same way. Artificial Analysis scores Spark Open at 39 on its Intelligence Index, the highest score it gives any open weight model built outside China, with seven Chinese models ahead of it and DeepSeek level with it at 39. That is also the honest way to read Meta’s headline claim. “Most capable United States open weight model” is a carefully drawn circle: inside it, the claim looks defensible on Meta’s numbers. The global open weight race, on the evidence Meta itself published, is still led from China.

The fair one line verdict: a top tier agent and tool user, a mid pack coder among the giant open models, and a claim sheet that should be retested once independent runs land.

Get started

Frequently asked questions

Is Muse Spark Open free?

Yes. Meta released the weights on October 15, 2026 as a free download under the Muse Model License, and downloading them costs nothing. Running the model is where the money goes: either serious hardware of your own, or per token charges on a hosted API.

Can I run Muse Spark Open on my laptop?

No. Meta says the model needs about a terabyte of memory, and its own desktop demo links two Mac Studios with 512 gigabytes each. Community builds have reportedly squeezed it onto machines with about 256 gigabytes, which is still far beyond any laptop.

Is Muse Spark Open the same model as the one in the Muse app?

They are siblings, not twins. The Muse app runs Muse Spark, Meta’s closed model, while Spark Open is the open weight version Meta released for developers on October 15, 2026. Spark Open shares the Spark base and adds extra safety training, because it runs without Meta’s own system around it.

Can my business use Muse Spark Open commercially?

For most businesses, yes, free of charge under the Muse Model License, as long as your product shows the words “Built with Muse Spark Open”. The big exception is selling the model itself as a service: companies with more than $20 million in revenue must notify Meta and sign a separate agreement first. Using it inside your own company, or inside your own product, is exempt from that rule.

Where do I download Muse Spark Open?

The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta ships two builds: the full precision BF16 checkpoint at about 1.6 terabytes, and the compressed NVFP4 build at about 530 gigabytes. Most people skip the download entirely and call a hosted API instead.

Is Muse Spark Open available in India?

The model is; the app is not. Developers in India can download the weights and call the hosted APIs today, subject to each provider’s own country and payment rules. The Muse app itself is still limited to the United States and Canada, and this release does not change that.

Built with Muse is an independent community project. Not affiliated with or endorsed by Meta.