Meta Releases Muse Spark Open: What an 816B Open-Weight Model Means for You

Meta released the weights of Muse Spark Open on 15 October 2026. It is a mixture-of-experts model with 816 billion parameters, 42 billion of them active per token, and a one-million-token context window, and Meta pitches it at coding and agent work. Our take: this matters to developers and startups that want a frontier-class model they can host themselves. For everyday users it changes very little right now, and almost nobody will run it on their own machine.

What actually shipped

We went to the model card on Hugging Face and Meta’s research blog rather than the coverage. The key details:

• Inputs and outputs: it takes text, images, video and audio, but it only answers in text. It does not generate images or speech.
• Formats: full-precision BF16 and a compressed NVFP4 build. Community GGUF and MLX conversions appeared within days.
• Hardware: the card asks for at least 700 GB of GPU memory for NVFP4 and 2 TB for BF16. Meta’s showcase setup is two M5 Ultra Mac Studios linked together.
• Access: the Hugging Face download is gated behind a form asking for your full legal name, date of birth and organisation.

TechCrunch’s Lucas Ropek described it as the most capable US open-weight model. That claim rests on Meta’s own evaluations, and Meta’s own table still shows several Chinese open models ahead of it on the general index, even where it leads on some agent tests. We are not reproducing the scores, because none of them are independent.

What it costs

The weights are free. In practice most people will rent it. When we checked on 17 October, OpenRouter listed it at $1.25 per million input tokens and $4.25 per million output tokens, and Ollama‘s cloud price table shows the same figures. OpenRouter charges exactly the same for Meta’s closed Muse Spark 1.3, so the open release does not make tokens cheaper. What you gain is the right to keep the weights, host them where you like and fine-tune them.

The licence is generous, with three catches

We read the full Muse Model License. It permits commercial use, fine-tuning, distillation and redistribution with no fee. Three conditions stand out:

• Attribution in the product. You must prominently display “Built with Muse Spark Open” in the UI, documentation, website or settings of anything that uses it.
• A ceiling for API resellers. If you sell access to the model itself as a service and your group’s revenue passes $20 million in any 12 months, you need a separate agreement with Meta. Internal use and features inside your own product are excluded.
• Adults only. The usage policy says the model is not intended for under-18s, which rules it out for many education products.

The Arabic question nobody has answered yet

The model card says only that it is multilingual and was trained on data from more than 100 languages. We found no mention of Arabic and no Arabic results, and the only language tag on the Hugging Face page is English. Quality in Modern Standard Arabic, let alone Gulf, Levantine, Egyptian or Maghrebi dialects, is unknown. For comparison, the Arabic-first Jais 3 family shipped the same week under a plain Apache 2.0 licence, and its smaller models run on ordinary hardware.

If you serve Arabic-speaking users, run a quick test before committing. Take a formal email in MSA, a WhatsApp message in your dialect and a scanned Arabic document with a table. Send each one to Muse Spark Open through OpenRouter and to your current model, and ask both for a summary and a suggested reply. Watch for English words leaking into the Arabic and for tables read in the wrong direction. It will cost you a few cents and tell you more than any leaderboard.

What to do about it

• Everyday users: nothing yet. It will not run on a laptop and it does not change the apps you use.
• Developers and startups: try it through OpenRouter or Ollama Cloud on agent and coding tasks, and plan for the attribution line before you ship.
• Teams with their own GPUs: this is one of the few models at this level you are allowed to host yourself, which matters where data-residency rules keep data inside the country. The weights are on Hugging Face.

Sources

• Meta research blog: Muse Spark Open announcement
• Model card and licence text on Hugging Face
• TechCrunch coverage
• Ollama Cloud price table