Meta Muse Spark Open, Explained in Plain English

Most AI news this month has been about who can keep a model the most locked up. Meta just did the opposite, on purpose, and at a scale most companies can't match.

On October 15, 2026, Meta released Muse Spark Open, an open-weight AI model that anyone can download, inspect and run on their own hardware, provided they can find the hardware for it. Meta calls it the most capable US open-weight model to date, and for once, that kind of claim is worth a closer look rather than a shrug. If you've used Meta Muse, the company's AI agent, this is a different thing entirely: Muse Spark Open is the underlying model, not the consumer app.

Here's what actually got released, what "open" means this time, and what it means for anyone who isn't going to download an 800-billion-parameter model onto their laptop. If you want to see it next to Google's and OpenAI's launches from the same fortnight, our Gemini 4 Argon vs GPT-6 vs Muse Spark Open comparison puts all three side by side.

What is Meta Muse Spark Open?

Muse Spark Open is a large language model built on a mixture-of-experts design, with 816 billion parameters in total and 42 billion active for any single response. Meta's release post describes it as natively multimodal, with a context window of 1 million tokens, and built for coding and agentic work. The Hugging Face model card says it takes text, images, video and audio as input and replies in text.

A context window of 1 million tokens is enough to hold an entire book, or a very long codebase, in a single conversation. Meta's PyTorch post says the model is trained for agentic coding: multi-file refactors, debugging sessions and long-running tasks, with reliable tool calling.

The weights, the actual numbers that make the model work, are published in two formats, BF16 and NVFP4. Meta says they are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. If you'd rather use it through an API instead of hosting it yourself, Meta lists hosted providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

Where it fits in the Muse family

Muse Spark Open did not appear from nowhere. Meta's Muse family started in April 2026 with a closed model called Muse Spark, which Meta kept proprietary. The PyTorch post says the open version builds on Meta's August release of Muse Spark and shares the same base model as Muse Spark 1.3, with additional safety post-training for use without system guardrails. Meta has also released Muse Glimmer, a small dense model. The consumer assistant built on this family is covered in our Meta Muse AI agent explainer.

Meta had said an open-weights Muse Spark release was planned, and this release delivers it. Reports quote Alexandr Wang, who has announced Muse releases for Meta, as saying the company will generally not open-source its most powerful models, which means Muse Spark Open is not the strongest model in Meta's lineup. That matters when you read the phrase "most capable". It describes the best open-weight model from a US lab, not the best model Meta has. For how it stacks up against Google's and OpenAI's newest releases on price and access, see our three-way model comparison.

What "open" really means here

"Open" is doing a lot of work in that name, so it's worth being precise about what you're actually getting.

Meta is releasing this under something it calls the Muse Model License, not a standard open-source license like MIT or Apache. Meta's own post says the license has limited commercial terms and lets many developers use the model for free. The Hugging Face model card lays out the conditions. You may use, modify, fine-tune, distribute and sell the model, but products must prominently show "Built with Muse Spark Open". If a company offers model access to third parties as a service and its aggregate gross revenue passes $20 million in a 12-month period, it must notify Meta and sign a separate agreement first. Internal use, research and using the model as a component of your own product are excluded from that threshold. A usage policy also bans uses such as malware and weapons development.

Access on Hugging Face is gated. The card asks for a legal name, date of birth and organization before you can download. Critics have pointed out that public weights don't make the license itself open-source in the traditional sense, and these conditions show why. Public weights also don't automatically mean public training data, which is a separate debate that came up this month in OpenAI's decision to stop training on the live internet.

In practice, this means: read the license before you build a product on top of this, especially if your company is large or the product is commercial. Treat "open-weight" and "open-source" as two different things, because right now a lot of coverage is using them interchangeably when they aren't. JavaScript developers ran into the same confusion this month for a different reason: Deno's code staying open source didn't stop Cloudflare's acquisition of its team from raising the question of is Deno dead. "Open" describes the code, not the company's plans for it, in either case.

Can you actually run it?

Technically, yes. Practically, almost nobody can run the full model at home.

The model card says the model needs a GPU setup with at least 700 gigabytes of VRAM for the NVFP4 version and 2 terabytes for BF16, before you've asked it a single question. Meta's own suggestion for running it locally is two linked Apple M5 Ultra Mac Studios. Meta says each machine's integrated GPU can use up to 512 GB of unified memory, so two of them can hold the full model. That tells you everything about who this release is really for: researchers, cloud providers and well-funded labs, not someone with a gaming PC.

For almost everyone else, the realistic way to use Muse Spark Open is through one of the API providers already hosting it, the same way most people use GPT-6 or Gemini without owning the hardware behind either one. The value of "open weights" at this scale isn't that you'll run it in your bedroom; it's that researchers can inspect how it works, and competitors aren't locked out of building on it the way they would be with a fully closed model.

If you do plan to pay for hosted access, price your expected volume first. The free Claude vs GPT vs Gemini API cost calculator runs that math in your browser, and our Gemini API price increase report shows how quickly a bill can change.

How good is it, really?

Meta's own benchmarks naturally make Muse Spark Open look strong, which is true of every company announcing its own model. On its Hugging Face table, Meta reports 39 on the Artificial Analysis Intelligence Index, below Kimi K3 at 45 and GLM 5.3 at 44, plus scores such as 89.0 on GPQA Diamond and 76.4 on Terminal-Bench 2.1. Those are self-reported, with the highest reasoning effort.

Independent testing tells a more mixed story, and the numbers depend on which leaderboard and which day you look at. On the Vals AI leaderboards, Muse Spark Open placed third on the Finance Agent v2 benchmark at 61.34% as of late October, behind Gemini 4 Argon, while other Vals pages for the same model show lower ranks, including 46.61% on one version of the overall Vals Index. Different versions weigh tasks differently, so treat any single rank as a snapshot. The honest summary is that it looks strong on some agent tasks and mid-pack on others, rather than a clean win across the board. That's a different picture from "most capable open-weight model", and it's worth remembering whenever a company announces a model using its own numbers.

If you want a wider sense of how fast this field moves and why every announcement this month claims a new record, how AI actually works is a good plain-English starting point before you take any single benchmark too seriously.

What developers are actually doing with it

In the first days after release, the real activity isn't happening on laptops, it's happening through the API providers now hosting the model. Seven providers were listed at launch, which usually signals enough developer interest to justify the integration work.

The typical use for a model like this isn't running it as-is for everyday chat. It's one of two things. The first is fine-tuning it on a narrow, specific dataset a company already owns, where an open model can be adjusted in ways a closed one never allows. Meta's PyTorch post says its TorchTitan training platform supports Muse Spark Open out of the box for supervised fine-tuning, LoRA and reinforcement learning. The second is using it as a teacher model to train a much smaller one through distillation, which compresses what the big model learned into something that fits on ordinary hardware. The license explicitly lists distillation among permitted uses.

Because the model is trained for agentic coding, it also competes for a spot in coding tools. If you are curious how far describing an app in plain words has come, our report on vibe coding Android apps shows the same idea on the consumer side.

Safety and the open-weight question

Open weights bring a specific worry: once a model is public, its maker cannot take it back or add guardrails later. Meta says Muse Spark Open went through additional safety post-training to make it more robust against misuse without system guardrails, and that it met the release thresholds in the Meta Superintelligence Scaling Framework. Meta's post says a full evaluation report would be available soon, so the detailed evidence was not public when the model launched.

That gap is part of a larger debate about who can switch AI systems off or limit them, which we covered in does AI have a kill switch. Open weights make that question harder, because there is no single switch to flip.

Why Meta is doing this now

Meta's own post frames the release as part of bringing advanced AI to everyone and says openness comes with extra responsibility for safety. It does not mention China. News agency coverage, though, described the release as a rival to Chinese open-weight models, several of which have gained real traction with developers because they're free to inspect and modify. That reading is plausible, and it is the press's framing, not Meta's. Chinese models are also reaching the products of US companies: Apple's first Apple Intelligence features for China will include Alibaba's Qwen, which we cover in our Apple Intelligence China release date report.

There's a business logic underneath the generosity, too. A developer who builds their product on Meta's model, under Meta's license, inside Meta's tooling, may find it harder to move to a competitor later. That is one reasonable reading, not something Meta has said. Open weights can be a genuine gift to the research community and a long-term strategic move at the same time. Neither reading cancels the other out.

What this means if you're not a researcher

If you're not fine-tuning foundation models for a living, the direct impact on you is small today. The indirect impact is bigger than it looks: models like this one get distilled, wrapped and repackaged into smaller tools within months, the same way earlier open releases fed into products people actually use day to day. It's part of a broader pattern of AI agents and models moving into ordinary life, usually a few steps removed from the original announcement.

It's also worth watching simply because of who's involved. Meta already controls Facebook's feed through tools like Facebook Your Algorithm, and a company building both the consumer-facing AI and the underlying model stack is accumulating a lot of influence over how AI reaches ordinary users. Our look at Meta Muse reaching 10 million downloads shows how quickly that consumer side can grow. Whether that's good or concerning depends entirely on what gets built with releases like this one.

An agentic model that can browse the web and use a computer also raises a practical question for anyone who lets AI tools touch their files, which is why our guide to macOS Full Disk Access for AI agents is worth reading before you give any agent broad permissions.

If you're a developer who does plan to build against one of the API providers hosting Muse Spark Open, a JSON to TypeScript converter turns a sample API response into clean types in seconds, which is a small thing that saves real time once you're actually wiring this into a project.

A short checklist before you build on it

Before you commit a project to any open-weight model, ask a few questions. Does the license allow your use, including the revenue threshold and the attribution notice? Do you have the hardware, or a hosted provider whose price you have tested at your real volume? Have you tried the model on your own task rather than relying on a leaderboard? What is your plan if the license, the provider list or the price changes? And where will your data go if you use a hosted API? Writing the answers down takes ten minutes and can save weeks of rework. The same habit of testing a tool on your own task, not on a launch-day demo, is the heart of why most people use AI wrong.

The bottom line

Meta Muse Spark Open is a genuinely large, genuinely capable model that almost nobody will run themselves, released under a license that's open but not open-source, with independent benchmarks that are solid in places and unremarkable in others. It's a meaningful move in the open-weight race, and also a reminder that the word "open" on an AI announcement is always worth a second look before you repeat the headline.

For most people, the practical takeaway is simple: the tools built on top of models like this will matter far more than the model itself ever will.

FAQs

What is Meta Muse Spark Open?

It's an open-weight AI model Meta released on October 15, 2026, with 816 billion total parameters, 42 billion active parameters and a context window of 1 million tokens. It accepts text, images, video and audio, replies in text, and is built for coding and agentic work.

Is Muse Spark Open free to use?

The weights are free to download under the Muse Model License, which allows many developers to use it, including commercially, with conditions: a "Built with Muse Spark Open" notice, and a separate agreement for companies that offer the model as a service and pass $20 million in revenue in a 12-month period. It isn't a standard open-source license, so read the terms before building a commercial product on it.

Can I run Muse Spark Open on my own computer?

Not realistically. The model card says it needs at least 700 GB of VRAM for the NVFP4 version and 2 TB for BF16. Meta suggests two linked Apple M5 Ultra Mac Studios; most people access it through an API provider instead.

How does Muse Spark Open compare to ChatGPT or Gemini?

Independent results are uneven and depend on the leaderboard version. It placed third on the Vals AI Finance Agent v2 benchmark in late October, behind Gemini 4 Argon, but ranks lower on other Vals pages. It's strong in specific areas rather than a clear all-round leader. Our model comparison has the side-by-side view.

Is this the same as the Meta Muse app?

No. Meta Muse is the company's consumer AI agent that books, shops and manages tasks. Muse Spark Open is the underlying language model Meta is releasing for developers and researchers to build on.

Who hosts Muse Spark Open through an API?

Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Provider lists and prices change, so check each provider before you build, and run your expected volume through the API cost calculator.