Meta released the weights of Muse Spark Open on 15 October 2026 and described running the model across linked M5 Ultra Mac Studio machines using MLX. This is a separate release from the previously introduced Muse Wonder. It expands the options for operating a large model locally, but desktop deployment does not mean an ordinary household computer.
Multimodal input, text output
The model card lists 816 billion total and approximately 42 billion active parameters, a one-million-token context, and text, image, video and audio input. Output is text. Understanding a video therefore is not the same capability as generating a video. Actions inside another application are performed through the surrounding environment and should be distinguished from the model's output modality.
Baseten confirms availability through its Model APIs and Dedicated Inference in NVFP4. The provider describes compressed MoE blocks and a mixture of local and global attention as ways to reduce inter-GPU traffic and long-context memory. These explanations come from a deployment participant, rather than an independent SIExpo measurement.
Our interpretation is that multimodality changes the evidence supplied with a task. For a hypothetical interface bug, a text model receives a person's description; a multimodal model can also receive an image of the problem. That does not remove the need for a real test after an edit. Correctly identifying a visual mismatch and correctly changing the code remain separate outcomes, even when one agent performs both steps.
Memory remains a substantial condition
Meta describes two machines with up to 512 GB of unified memory each. The model card gives separate GPU-serving requirements of at least 700 GB VRAM for NVFP4 and 2 TB for BF16. These are developer-published requirements, not a certification that arbitrary hardware with a similar total memory will behave identically. We have not verified local hardware availability or pricing.
Adding memory capacities does not describe communication speed between devices. For a hypothetical internal deployment, we would examine how the model is distributed and how it responds at the expected conversation count. Full context, concurrent requests and precision affect that operational choice. Completing one request is a different test from serving an entire team throughout its working day.
MLX LM is a public toolkit for language models on Apple silicon. Its availability should not become an assumption that every toolkit version supports every new release equally. A reproducible pilot would record the exact model, conversion, software version and settings, rather than retaining only a general product name in a deployment note. Those details matter when an update changes behaviour.
Downloadable weights have a specific license
Muse Spark Open uses the Muse Model License with commercial conditions. We do not describe it as Apache-licensed or as unconditional permission for every business use. An adopting team needs to read the applicable text against its application. Being able to download weights does not by itself answer questions about redistribution, providing a service or associated obligations.
Vendor quality evaluations are also separate from availability. For a meaningful internal assessment, we would compare the same tasks and record elapsed time, human corrections and final results. If one configuration uses more reasoning or a different agent, that difference needs to remain visible. Otherwise, an apparent model comparison may partly describe a difference in the surrounding workflow.
SIExpo has not reproduced Meta's evaluations. Muse Spark Open's release is a concrete event, while its suitability for a particular team needs evidence from that team's own work. The useful decision combines access, licensing, infrastructure and completed tasks; none of those questions can be settled by the word open alone.
Editorially checked: 2026-10-16T16:26:59.575073+00:00 (UTC). Cover: conceptual AI illustration.
