Meta's 816B Open Model: What You Actually Get

Meta published the weights for Muse Spark Open on October 15, calling it the most capable US open-weight model to date. It is an 816-billion-parameter multimodal model with a million-token context window, released under a licence Meta describes as carrying limited commercial terms. Both halves of that sentence matter, and the second half is the one that decides what it means for your business.

What Meta shipped

Detail	Per Meta
Size	816B total parameters, about 42B active per token
Inputs and context	Text, image, video and audio; 1M-token context
Trained for	Long-horizon agentic work, including coding and support agents
Where to use it	Seven API platforms at launch, plus four cloud model catalogues
Desktop option	Two Apple M5 Ultra Mac Studios running MLX
Licence	Muse Model License, described as having limited commercial terms

The distribution is the genuinely notable part. Meta lists availability through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with weights downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Day-one support across the major inference and fine-tuning frameworks came with it.

The licence is the clause to read

Meta's own wording is that the weights are released under the Muse Model License with limited commercial terms, which allow many developers to use the model for free. Many is not all, and limited is a real qualifier rather than a hedge.

Custom model licences commonly attach conditions of a few recognisable kinds: thresholds above which different terms apply, restrictions on building competing products, and attribution requirements. Which of those appear here, and at what levels, is a question for the licence text against your specific plan.

Threshold conditions deserve the most attention because they are invisible while a project is small and decisive once it works. Checking at pilot stage costs twenty minutes; discovering it after you have built a product on top is a different conversation. The general version of this check is in what to look for in AI terms of service.

You are probably not hosting this

The desktop headline is a real engineering achievement and a poor planning assumption. Meta says Muse Spark Open is the first model in its class to run on a desktop, which turns out to mean two Apple M5 Ultra Mac Studios with the MLX framework.

That is a hardware purchase, a person who can operate it, and an ongoing maintenance commitment, for a single machine's worth of throughput with no redundancy. For a business without an infrastructure team, the arithmetic rarely favours it over paying per token, and the smaller models that genuinely do run on ordinary hardware are covered in running LLMs locally in Canada.

Which leads to the correction most worth making. If you reach this model through a hosted API, your data goes to that host under that host's terms, exactly as with a closed model. The privacy and residency benefits people associate with open weights come from self-hosting, not from the licence, so data residency questions need answering about the host rather than the model.

What it genuinely buys you

Provider choice, immediately. Seven API platforms serving the same weights at launch is price competition on an identical product, which does not exist for a closed model. If one host raises prices or has an outage, the others run the same thing, and switching is a configuration change rather than a migration.

Fine-tuning on your own data. Post-training support shipped at release across standard frameworks and through managed services from some of the hosts. For businesses with a specialised vocabulary or a narrow repeated task, that is the path to a model that fits the work rather than one you prompt around.

An exit that exists. The option to move hosts, or eventually to self-host, has value even unexercised, because it changes what a renewal conversation sounds like. That is the practical core of model hedging to avoid vendor lock-in.

Who should look this quarter

Businesses already paying meaningfully for a frontier model on a well-defined task, where running the same prompts through a cheaper host is a day of work and the comparison is straightforward. Businesses with a fine-tuning case, meaning a repeated task with a lot of their own examples. And businesses in sectors where a credible self-hosting path has to exist on paper, even if the near-term deployment is hosted, which is the situation described in self-hosted models and compliance.

For everyone else this is a market development rather than a decision. Open weights at this capability level pushes prices down across the board, including on the closed models you are already using, and that benefit arrives without you doing anything.

Frequently Asked Questions

What is Muse Spark Open?

An open-weight model Meta released on October 15, 2026, which the company describes as the most capable US open-weight model to date. It is a sparse mixture-of-experts model with 816 billion total parameters and roughly 42 billion active per token, takes text, image, video and audio input, holds a 1 million token context window, and is trained for long-horizon agentic work such as coding and customer service agents.

Is Muse Spark Open actually open source?

The weights are published, which is a real and useful thing, and the licence is Meta’s own rather than a recognised open source licence. Meta describes it as the Muse Model License with limited commercial terms, which it says allows many developers to use the model for free. Open weights and open source are different claims, and the only way to know whether your intended use is permitted is to read the licence against that specific use.

Can my business self-host this model?

Almost certainly not, and the desktop headline is doing some work. Meta says it is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with the MLX framework. That is an impressive engineering result and it is not a deployment plan for a company that does not have a hardware budget and someone to run it. Most businesses will reach this model through the same kind of API they would use for a closed one.

So what does open weights buy a business?

Three things that matter commercially. Provider choice, since the model launched across seven API platforms plus four cloud model catalogues, which means price competition and no single vendor controlling your access. The ability to fine-tune it on your own data with tooling that is already supported. And an exit path, because the option to move or self-host later exists in a way it does not with a closed model, even if you never exercise it.

Does using an open-weight model keep our data private?

Only if you run it yourself or contract for that specifically. Accessing an open-weight model through a hosted API means your data goes to that host under that host’s terms, exactly as it would with a closed model. The privacy benefit people associate with open weights comes from self-hosting rather than from the licence, so if data residency is your reason for looking, check where the inference actually happens before assuming anything changed.