Meta Releases Muse Spark Open, a Large Multimodal Model for Local Inference on Macs

On October 15, Meta released Muse Spark Open, a natively multimodal Mixture of Experts (MoE) model whose weights are available through the Hugging Face platform. The model has 816 billion total parameters, of which 42 billion are active during inference, and supports a context window with a capacity of one million tokens.

The release enables users to deploy and modify the model themselves outside hosted APIs. However, this does not mean it will be easy to run on a typical personal computer. For local deployment of the full version on Apple silicon, Meta and Apple specify a configuration of two Mac Studio computers with the M5 Ultra chip and 512 GB of unified memory in each device.

Muse Spark Open Uses an MoE Architecture

The MoE architecture divides a model into multiple specialized components, activating only some of its parameters when processing a single input. For Muse Spark Open, Meta therefore lists 42 billion active parameters despite the model’s total size of 816 billion parameters.

The model is natively multimodal. Its model card on Hugging Face lists multimodal inputs and a context window of one million tokens. This is substantially larger than the context windows of many models used in local tools, but it also increases memory and inference-system requirements.

Meta describes Muse Spark Open as its “most capable American open-weight model.” This is the manufacturer’s assessment based on its own benchmarks, not an independently confirmed comparison with competing models.

Two Mac Studio Computers and Distributed Inference

The release’s most notable technical feature is its optimization for distributed execution on Apple devices. Meta specifies a configuration of two Mac Studio computers with M5 Ultra, with each device equipped with 512 GB of unified memory. The connection is intended to use RDMA over Thunderbolt 5 and the MLX framework.

In its own research paper, Apple published a procedure for distributed execution of a quantized version of the model in the MLX framework across multiple Mac Studio computers. Quantization reduces the model’s memory requirements by using lower-precision data, but the specific result depends on the selected version and inference software.

The requirement for two fully equipped workstations also defines the practical limits of this scenario. According to the stated configuration, full local operation is not intended for a single typical Mac. Thus, while the model demonstrates distributed inference on Apple silicon, it does not represent a readily available local solution for most developers.

The Weights Are Available, but the License Has Conditions

The Muse Spark Open weights are published through Hugging Face under the Muse Model License. It permits broad use but includes conditions and limited commercial provisions. In this case, the open-weight designation means that the model weights are available; it cannot automatically be equated with a fully open-source project without licensing restrictions.

For developers, the availability of the weights is especially important for self-hosting, modification, and experimentation without having to send every input to a remote API. Practical usability will depend on the available hardware, the quantization used, and support in local tools.

What to Watch After the Release

After the launch, it will be important to monitor independent measurements of quality, speed, and memory requirements, both in MLX and on GPUs from other manufacturers. Meta also says it will provide a full evaluation report and additional safety materials. The availability of more practical quantized variants and their integration into local inference tools will also be important.

Verified and updated: 10/16/2026 06:19