Meta opens new AI model, but with special terms for large companies

On October 15, 2026, Meta released the language model Muse Spark Open as a so-called open-weight model. That means the model's inner structure, known as its weights, can be downloaded and used by others, rather than the model being usable only at Meta.

Meta itself describes the model as the most powerful American open-weight model to date.

What can the model do?

Muse Spark Open has 816 billion parameters in total, but uses 42 billion active at a time. Parameters are the settings a model has learned during training. The model can work with more than text and has a context window of 1 million tokens. Roughly speaking, that corresponds to how much text it can keep track of at once. The focus is reasoning, tool use, collaboration among multiple AI agents and software development.

The model can be downloaded from Hugging Face and is available from Amazon, Google and Microsoft. Access is also possible through Groq, OpenRouter and Together AI, and it can run locally on a Mac Studio with M5 Ultra. At Together AI, it costs $1.25 per million input tokens and $4.25 per million output tokens.

The license is not fully open

The model is released under Meta's own Muse Model License, which is not a classic open source license. Use is free for many developers, but there are terms. Products must display the text "Built with Muse Spark Open." Companies with more than $20 million in annual revenue must notify Meta and enter into a separate agreement. Research and internal use are exempt. According to Let's Data Science, downloading also requires information about identity and organization.

The figures come from Meta

In Meta's own comparison, Muse Spark Open scores 39 on the AA Intelligence Index versus 45 for GLM 5.3. On tool use, it gets 89.2 versus 84.2 for GLM 5.3. On terminal coding, it gets 76.4 versus 85.0 for Kimi K3, and on DeepSWE 60 versus 68 for Kimi K3. The picture is thus mixed. Nearly all the figures are Meta's own, and independent verification was lacking at launch.

Other open models

Meta's table does not include DeepSeek or Reflections Beam, and the sources do not compare the three directly. DeepSeek V4.1 came out in September 2026 under the MIT license, which allows commercial use. According to a model database, it has 552 billion parameters and 1 million tokens of context, but the parameter count is unclear, as other summaries mention 1.6 trillion. Its Flash variant typically costs $0.30 and $1.20 per million input and output tokens.

Reflections Beam 1.0 has 501 billion parameters, of which 23 billion are active, and a context window of around 1 million tokens. On OpenRouter, it costs $0.25-0.40 for input and $1.00-1.20 for output per million tokens. According to OpenRouter, it is not available via EU regional routing, which could matter for European companies.