Muse Spark Open: Meta's New Open-Weight Model, Explained

Meta released the weights of Muse Spark Open on October 15, 2026, calling it the most capable open-weight model from a US lab. Anyone can request the files on Hugging Face and run them on their own hardware, or use the model through hosted services, under a license that is free for most businesses but comes with conditions. It is large: 816 billion parameters, 42 billion of them active for each token it produces, with a 1-million-token context window.

What "open weights" means, in one minute

A model's weights are the trained file that holds everything it learned. With closed models such as Claude or GPT-6, you can only rent access through the company's apps and API. With open weights you can download that file, run it on your own servers, adjust it to your data, and never send a prompt to the original maker.

It is not the same as fully open source: Meta's model card describes where the training data came from, but the data itself is not published. The launch post lists where you can get the weights.

The local option is the headline for Meta. It says two M5 Ultra Mac Studios, each with up to 512 GB of unified memory, can hold the whole model and run agent tasks entirely on-device. That puts a large modern model within reach of an office, not just a data center.

How good is it?

Good, but "most capable US open-weight model" is a narrower claim than it sounds. The strongest open models today come from Chinese labs, and Meta's own comparison table on the model card shows three of them ahead on the Artificial Analysis Intelligence Index, an independent score that averages many tests.

Where Meta does claim a clear lead is safety. On its internal cyber misuse test, Meta reports that Muse Spark Open refused 99.2% of jailbreak attempts, against 75.5% for Kimi K3 in the same setup. Those are Meta's numbers, but for a business deploying a model to customers, refusal behavior matters as much as raw scores.

You can compare the wider field on our AI model leaderboard, where Muse Spark Open is too new to appear on the arena boards yet.

The license, in plain English

The Muse Model License is unusually short. It grants free use, including changing, fine-tuning and selling products built on the model, subject to a few conditions:

Should a small business care?

For most owners, an open-weight model matters in two situations. The first is privacy: if client contracts or regulations stop you from sending data to an outside AI company, a model you host yourself, or in your own cloud account, can solve that. The second is volume: if you run thousands of automated tasks a day, a hosted open model can cost less than the top closed models.

If neither applies, a closed model in a normal app is still simpler.

Before switching anything, run the same real task through your current model and Muse Spark Open, for example via OpenRouter, and compare:

You are helping [YOUR BUSINESS]. Here is a real task from my week: [PASTE THE TASK AND ANY TEXT IT NEEDS].
Do the task completely. Then list any facts you were unsure about and any assumptions you made.
Keep the answer under [NUMBER] words and use the same format I would send to a client.

Score both answers on accuracy, tone and how much editing they need. Cost per task, not per token, is what decides it. Our guide to AI automation for small business shows how to pick the tasks worth testing.

What to watch next

• The full safety report. Meta says a complete evaluation report is "available soon"; until then, the model card is the only detailed source.
• Independent votes. Arena rankings, where people compare answers blind, will show whether real users prefer it over Kimi K3 and GLM-5.3.
• Hosted prices. As more providers serve the model, per-token prices tend to shift. Check two or three hosts before you budget a project around it.