Meta releases the weights of Muse Spark Open, a model with 816 billion parameters

Meta released the weights of Muse Spark Open on October 15, 2026, which means the model can be downloaded and run outside Meta's own services. It is a mixture-of-experts model: a routing layer sends each token to a small set of specialized subnetworks instead of through the whole network. Of its 816 billion parameters, about 42 billion are active per token. The model reads text, images, audio, and video, answers in text, and has a context window of 1,048,576 tokens. Meta says it built the model for coding and agent work, and calls it the most capable US open-weight model to date.

What it takes to run at home

All 816 billion parameters must sit in memory, even though only a fraction of them works on each token. Meta's desktop setup therefore uses two Mac Studio computers with the M5 Ultra chip, each with 512 GB of unified memory, running Apple's MLX framework. Meta says this is the first model in its class that runs on a desktop this way.

Hugging Face tested a single machine. Quantization stores the weights with fewer bits, so the model needs less memory. According to Hugging Face, a 3-bit version of the model took about 344 GB on an M3 Ultra with 512 GB and generated about 21 tokens per second. It also notes that even the 3-bit version needs more than 256 GB of memory.

Data center requirements are higher. The full-precision BF16 checkpoint, the saved file with the model's weights, is about 1.5 TB, and Hugging Face runs it in Transformers across four nodes of eight Nvidia H200 GPUs each. A second checkpoint in NVFP4, a 4-bit format for Nvidia's Blackwell GPUs, runs on a single node with B200 GPUs.

Where to get it and what it costs

Hosted access started on release day through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. On October 15, 2026, every provider listed on OpenRouter charged $1.25 per million input tokens and $4.25 per million output tokens.

You can download the weights from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. On Hugging Face the download is gated: you accept the license first, and access is granted after that.

The model runs in SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Reasoning cannot be switched off. There are four effort levels, from low to xhigh, and the default is the highest, so Hugging Face recommends low for simple tasks to spend fewer tokens.

What the license allows

Meta publishes the weights under its own Muse Model License rather than a standard open-source license. You may use, modify, fine-tune, and sell the model and models derived from it without paying Meta. Every product built on it must prominently display the line “Built with Muse Spark Open” in its interface, documentation, or settings.

One condition is aimed at large companies. If a company and its affiliates earn more than $20 million in any 12-month period, they must notify Meta and sign a separate agreement before selling access to the model as a service. Internal use, research, and features inside a company's own product are exempt. The usage policy also says the model is not intended for people under 18.

What has not been checked yet

The claim that this is the most capable US open-weight model rests on Meta's own evaluations, and no independent measurement had been published as of October 15, 2026. Meta says the model met the release thresholds of its safety framework and that a full evaluation report will follow.