Quick answer
Meta released Muse Spark Open on October 15, 2026. The open-weight model has 816 billion total parameters, 42 billion active per token, accepts text, image, audio, and video inputs, and has a 1,048,576-token context window. Its weights are downloadable and hosted inference is available through listed partners, but local serving requires unusually large memory capacity. The Muse Model License also includes use-policy, attribution, and a specific condition for large model-as-a-service businesses. Muse Spark Open is not listed in Chat AI's public model directory, so availability there is unverified.
Muse Spark Open is a downloadable model, not simply a new chatbot option
Meta released the weights for Muse Spark Open on October 15. Meta lists 816 billion total parameters and 42 billion active parameters in a sparse mixture-of-experts design, with a 1,048,576-token context window. The model accepts text, images, audio, and video, and returns text. These are release specifications, not an independent test of quality or speed.
Sources: Meta AI Research, Meta Model API
Open weights still come with a model license and usage rules
Meta calls the release open-weight and publishes it under the Muse Model License with an accompanying Acceptable Use Policy. The license requires attribution in products or documentation that use the model and sets a specific requirement for a provider whose aggregate revenue exceeds $20 million over a consecutive 12-month period and that offers the model or a derivative as a model-as-a-service product: it must notify Meta and enter a separate written agreement first. This is a high-level summary, not legal advice; read the license and policy for the exact terms that apply to your use.
Sources: Meta AI Research, Meta Superintelligence Labs on Hugging Face
The active parameter count does not describe the memory needed to serve it
Although only 42 billion parameters are active per token, Meta's technical documentation says serving still loads the full 816-billion-parameter weights. It lists an approximately 1.65 TB BF16 checkpoint and about 537 GB for the NVFP4 format; the documentation estimates roughly 2 TB of VRAM for BF16 after allowing for serving overhead. Meta's listed MLX configurations include two Mac Studio computers with M5 Ultra and 512 GB of unified memory each, or four with 256 GB each. These are specific high-memory setups, not a claim that the model will run on an ordinary laptop or single consumer desktop.
Sources: Meta Model API, Meta AI Research
Choose between self-hosting and a hosted inference provider
Meta documents two routes: download and serve a checkpoint on your own infrastructure, or call the model through a hosted inference partner without provisioning GPUs. Self-hosting gives an organization more control over where its inference runs, but requires suitable hardware, runtime setup, and its own safeguards. Hosted access lowers the infrastructure burden but means evaluating the provider's service, costs, and data terms. Muse Spark Open is not listed by exact name in Chat AI's public model directory, so its availability in Chat AI is unverified; check the current catalog rather than assuming a hosted model release appears in every chat app.
Sources: Meta Model API, Chat AI
Frequently asked questions
What readers usually ask
Does open-weight mean Muse Spark Open is open source?
Meta describes it as an open-weight release, but distributes it under the Muse Model License and an Acceptable Use Policy. Review those terms instead of assuming an unrestricted or standard permissive open-source license.
Can I run Muse Spark Open on one regular desktop computer?
Meta lists MLX setups using either two Mac Studio computers with M5 Ultra and 512 GB of unified memory each, or four with 256 GB each. Its general serving documentation lists much larger memory requirements than a typical personal computer, so check the exact checkpoint and supported hardware before downloading.
Can I use Muse Spark Open through a hosted API instead of running it locally?
Yes. Meta lists hosted inference partners as well as model catalogs where weights can be obtained. Hosted API availability and pricing depend on the provider, so check that provider's current terms.
Is Muse Spark Open available in Chat AI?
Its availability is unverified. Muse Spark Open is not listed by exact name in Chat AI's public model directory, so check the current directory or in-app selector for verified access.
Evidence