Meta released Muse Spark Open: what you can run, and what it takes

Meta released Muse Spark Open, an AI model you can download and run on your own hardware. It handles text, images, video, and audio as input, and produces text. Meta built it for coding and agents that work through tasks over many steps.

The hardware requirement gives you an idea of its size. Meta demonstrated it running locally on two Apple M5 Ultra Mac Studios, each with 512 GB of unified memory.

That’s a terabyte of memory between two computers.

I haven’t tested the model myself. This post is based on Meta’s October 15 announcement and its official model card.

What Meta released

Muse Spark Open has 816 billion total parameters, with 42 billion active for each token. Its context window holds up to one million tokens.

If those numbers mean little to you, here’s the practical meaning.

Parameters are the learned numbers that make up the model. The model uses a mixture-of-experts architecture, which selects a subset of its internal networks for each token it processes. That reduces the computation needed compared with using the whole model every time.

But the other parameters still need to be stored somewhere. A model with 42 billion active parameters doesn’t have the memory requirements of a 42-billion-parameter model.

The context window is how much information the model can work with in a request, including the conversation and material you provide. A large context can help an agent work with a repository or long documents. It also consumes memory, so the maximum supported context is another thing to account for when sizing a deployment.

The released weights come in BF16 and NVFP4 formats. NVFP4 uses lower precision to reduce the space the model needs. Meta notes that quantization can produce some quality differences compared with full precision.

If you want more background on what downloading a model gives you, I covered that in my guide to open-weight models.

What it can do

Meta focuses on software engineering and tool use. An agent can ask the model what to do next, execute a command or edit a file, then send the result back for another step.

For example, fixing a bug can involve reading several files, changing the code, running tests, and correcting a failed attempt. The software around the model provides those tools and controls what it can access.

The multimodal input is useful here. An agent could inspect a screenshot alongside the source code when investigating a broken interface.

Meta’s local demonstrations include making changes to Ice Cubes, an open-source Mastodon client, and building the app. Another demonstration uses DaVinci Resolve to review footage and create vertical crops. These are Meta’s demonstrations; I haven’t reproduced them.

The model’s output is text. Support for video and audio input doesn’t mean this release generates videos or audio.

Running it locally takes serious hardware

Meta’s model card lists at least 700 GB of GPU memory for NVFP4, or 2 TB for BF16. These are deployment requirements from Meta, rather than a promise that every workload fits within the same amount of memory.

The Mac setup uses Apple’s unified memory, which the GPU can access, and the MLX framework. Meta says two M5 Ultra Mac Studios with 512 GB each can hold the model and run agent workflows locally.

I would be careful about reading “runs on your desktop” as “runs on the computer you already own.” This is a substantial hardware setup. Most laptops and ordinary desktop PCs are far below the stated memory requirements.

Fitting the model is only part of the decision, too. I’d want to see how long it takes to read a large prompt, how quickly it generates answers, and what happens as the conversation grows. An agent may make many model requests before it finishes a task.

The launch supports several inference engines, including vLLM, SGLang, llama.cpp, and Hugging Face Transformers, alongside MLX. That gives teams options for deploying it on their existing infrastructure, provided they have enough memory.

You can use a hosted provider

You can try the model without buying the hardware. Meta lists hosted access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

I’d start there to find out whether the model is useful for my work. Provider pricing and performance are separate questions from whether the weights are free to download.

Self-hosting gives you control over where inference happens and which checkpoint you keep running. Hosted access saves you from operating the machines. If your reason for using this release is keeping code inside your own infrastructure, sending it to a hosted provider changes that arrangement.

The official Hugging Face collection contains the released checkpoints. Meta’s developer guide covers deployment.

Read the license before building around it

Meta released the weights under the Muse Model License, with commercial conditions. Downloadable weights don’t automatically give you the permissions of an MIT or Apache-licensed project.

The license permits uses including modification, fine-tuning, and distribution, subject to its conditions and attached usage policy. It also requires a “Built with Muse Spark Open” notice in the product interface or an allowed accompanying location, such as documentation or the website.

There is a specific condition for businesses offering Model as a Service, where third parties receive model outputs through an API or similar interface. If the licensee and its affiliates exceed $20 million in aggregate gross revenue during any consecutive twelve-month period, the license requires notifying Meta and entering a separate written agreement before offering this model or its derivatives as part of that business.

That revenue figure covers the licensee and its affiliates. It isn’t described as revenue from this particular model alone.

The same clause explicitly excludes internal use, research, and commercial activity outside Model as a Service. It also excludes using the model to power features in your own product that isn’t a Model as a Service business. The other license conditions still apply.

So “commercial use is banned” would be misleading. Calling it unrestricted would be misleading too. Read the actual terms for the product you’re building.

How I read the benchmark claims

Meta calls it the most capable US open-weight model. That geographic qualifier limits the claim. It doesn’t say the model beats every open-weight model worldwide.

Meta’s own comparison shows a mixed picture. Muse Spark Open leads the listed models on MCP-Atlas, a tool-use benchmark. It trails GLM 5.3 and Kimi K3 on the Artificial Analysis Intelligence Index and DeepSWE coding evaluation.

The model card’s benchmark table includes the settings and comparison models. Meta evaluated Muse Spark Open with xhigh reasoning effort.

That setting matters. The model supports low, medium, high, and xhigh effort. More reasoning can improve an answer while taking longer. For a coding agent I use throughout the day, I’d compare completed tasks and time spent waiting at the settings I’d actually use.

I wouldn’t pick it from the headline alone, or dismiss it because a different model wins one coding benchmark.

How I would try it

I’d use a hosted endpoint first and give it a small, existing bug to fix. Something with a test I can run and a result I can check.

Then I’d try a UI task with a screenshot. That would help me see whether its visual input improves the changes it makes. I’d compare it with my current coding agent on the same work, including failed attempts and the time needed to review its edits.

I would keep file access limited and require approval for consequential actions. Meta’s model card recommends additional protections around the model when it can use tools. Running the weights yourself makes those controls your responsibility.

For my first experiment, the hosted providers in Meta’s announcement are enough. I’d want evidence from my own tasks before planning a local deployment with hundreds of gigabytes of memory.

Want me to talk about your product? You can sponsor this site.