Technical
Meta's Muse Spark Open Runs on Two M5 Ultra Mac Studios: What Local Costs
ยท RenderBob team
Meta's Muse Spark Open is an 816B language model with 42B active parameters. The documented local setup is two 512GB M5 Ultra Mac Studios, not two base machines.
Meta released Muse Spark Open on 15 October. It is an open-weight language model with 816 billion parameters in total, 42 billion active, and a 1-million-token context window. Meta says it runs locally across two M5 Ultra Mac Studios. It does not generate pictures or video.
What Meta published
Meta's research post calls the model natively multimodal and aims it at coding, reasoning, and multi-step agent work. It is optimised for the M5 Ultra with Apple's MLX framework. Meta repeats Apple's memory figures: up to 512GB of unified memory per Mac Studio, and up to 1.2TB/s of bandwidth, so two 512GB machines can hold the model. Meta's hardware page lists the Apple checkpoint as NVFP4, on two Studios with 512GB each or four with 256GB, linked over Thunderbolt. Apple's research note says four machines generate over 36 tokens per second. That figure is Apple's.
Weights are on Hugging Face, Microsoft Foundry Models, Google Cloud Model Garden, and Amazon SageMaker AI. Meta's named API hosts include Groq, OpenRouter, Together AI, Baseten, Fireworks, Modal, and Nebius. Supported engines include vLLM, llama.cpp, MLX, SGLang, and TensorRT-LLM. The weights ship under the Muse Model License, which Meta describes as limited commercial terms that still let many developers use the model free. Meta says the model met the release thresholds in its safety framework. The full evaluation report was described as still to come.
What the hardware actually is
Our 13 October Mac Studio piece put the M5 Ultra base price at $5,499. Retail coverage of that base configuration put it at 96GB, and Apple had listed the 512GB configuration for late October. Two base machines come to $10,998. That arithmetic is ours, and it is the wrong machine. Meta's documented setup is the 512GB pair, or four 256GB machines. We do not have a published price for the 512GB configuration. Open weights also do not mean unrestricted use. Read the licence.
Where it fits in a studio
- Scripts, logs, shot lists, and agent workflows that have to stay on machines you own. Picture generation is a different model.
- Meta's own example is the model driving DaVinci Resolve to review footage and crop selects. That is an agent on an editor, and it is Meta's claim.
- Keeping prompts and references on your own hardware is the practical answer when the material cannot be sent to a host.
- The same desks could hold an open image or video model later. Check for a maintained Apple-silicon build before you count on it.
Read the Muse Model License before you ship client work on it, and compare a hosted price against your real volume. RenderBob lets you choose local or cloud for each clip, so a private language model and a cloud video model can sit in the same cut.
What to watch
Independent speed and memory tests, how the commercial limits in the licence get read, and whether Apple-silicon builds appear for open video models.
More from the blog
- Fish Audio Drama 3: A 35B Open-Weight Voice Model You Direct in Plain Language
Fish Audio announced Drama 3 on 13 October. You write performance direction in square brackets, and one text can carry up to 16 speakers. The weights are on Hugging Face.
- Nano Banana 2.1: 14 References, Three Thinking Levels, and Targeted Edits for Storyboards
Google's Nano Banana 2.1 accepts up to 14 reference images and adds a medium thinking level. Artlist also describes edits that change a named area instead of redrawing the frame.