On October 15, 2026, Meta released the trained weights for Muse Spark Open, an AI model specialized for coding and work carried out by AI agents. It has 816 billion parameters in total, but uses about 42 billion during inference, and it supports contexts of up to 1 million tokens. Working with Apple, Meta has also published a configuration that runs the model on several M5 Ultra Mac Studios connected together, widening the options for operating large-scale AI in your own environment without depending on an external model API. However, even though the model can run locally, it requires large amounts of memory and fast connections between machines. When evaluating a deployment, it is important to distinguish between the number of parameters computed during inference and the memory capacity needed to hold the entire model.
Only 42 Billion Parameters Are Used for Inference, but the Memory Must Hold All the Weights
Muse Spark Open is a sparse Mixture of Experts (MoE) model, which selects the expert networks that handle each input. By narrowing the parameters used to process each token, it avoids having to compute all 816 billion parameters every time. However, according to Meta's developer documentation, the inference environment must load all of the model's weights. In other words, reducing the amount of computation does not reduce the required memory capacity by the same proportion.
The required capacity also varies greatly depending on the format in which the weights are stored. In BF16, the weights come to about 1.65TB, but in the NVFP4 format, which stores most of them in 4 bits and keeps some in BF16, they shrink to about 537GB. At runtime, however, memory is also needed for the KV cache, which holds the results of calculations on past tokens, as well as for intermediate computations and the execution software. Meta estimates that running inference in BF16 requires about 2TB of VRAM.
This additional memory consumption matters especially when AI agents are given long-running tasks. Even if the model supports up to 1 million tokens, fitting the weights in memory does not guarantee that contexts of that length or many concurrent tasks can be handled comfortably. The memory actually required depends on the length of the context and the number of tasks run in parallel.
Muse Spark Open accepts not only text but also images, audio, and video as input. Its output, however, is limited to text. It can be used for tasks such as fixing code while checking what is on the screen, but being able to accept audio as input is not the same as being able to generate audio.
"Muse Spark 1.3," which Meta announced on September 2, is offered mainly through Muse Code and the Meta Model API, and the release of its trained weights was described as a future plan. Muse Spark Open turns that plan into reality. Developers can now not only use the API but also obtain the model's weights and run inference in an environment they have set up themselves.
Memory-Saving Techniques That Support Long-Running Agent Work
Baseten, which provides inference services, describes compressed MoE and an attention structure that switches between a short window and the full context as the mechanisms that let Muse Spark Open handle long-running agent work. AI agents repeatedly run commands, check results, and revise code, so their history grows longer as the work proceeds. How much memory can be secured after the model weights are loaded therefore determines the performance of long-running processing and parallel execution.
In compressed MoE, a shared process reduces the dimensionality of the input vector before it is passed to each expert network. After the outputs of the experts are combined, the vector is restored to its original dimensionality. According to Baseten, this structure is designed to cut the number of parameters in the expert portion while also reducing the amount of data transferred between GPUs. Note that the router, which decides which experts to use, receives the uncompressed input.
Another technique is the combined use of local and global attention. Of the 62 layers, 46 use local attention, which looks at the most recent 2,048 tokens, and the remaining 16 use global attention, which looks at the entire context. With local attention, the cache for old tokens can be overwritten, so memory consumption can be kept lower than in a structure that holds the full context in every layer.
In a comparison published by Baseten, with a context of 128K tokens, a configuration that uses global attention in all layers requires 31GiB of KV cache per sequence. A configuration that combines local and global attention fits in 8.4GiB, cutting memory consumption to about one-3.7th. However, this reduction ratio is a comparison of the KV cache under specific conditions, and does not directly apply to the model's total memory requirements or processing speed. By combining weight compression through MoE with reduced history data through the attention structure, the model is made better suited to long-running agent work.
Comparing the Configurations Needed to Self-Host on Mac Studio and NVIDIA GPUs
Meta has published configurations for running Muse Spark Open in your own environment using Apple's Mac Studio and NVIDIA GPUs. For Mac, the assumed setup is either two 512GB-memory models or four 256GB models connected together. In NVIDIA environments, a single node with four Blackwell GPUs is the guideline for running the NVFP4 model. The table below summarizes the official hardware configurations and model weight sizes presented alongside the October 15 release.
| Environment | Format and weight size | Published configuration | Main software |
|---|---|---|---|
| Apple | NVFP4, about 537GB | M5 Ultra Mac Studio: two 512GB models or four 256GB models | MLX |
| NVIDIA | NVFP4, about 537GB | Single node with four Blackwell GPUs | SGLang, vLLM, TensorRT-LLM, etc. |
| NVIDIA | BF16, about 1.65TB | Configuration with eight GPUs such as B300 | SGLang, vLLM, etc. |
The weight sizes in the table do not include the KV cache and other overhead, and the memory actually required increases with context length, the number of parallel executions, and the software used. These are also example configurations for running the same model, and do not imply equivalent processing speed or deployment cost. Self-hosting on Macs likewise does not mean easily running the model on an ordinary laptop; it requires combining several machines with large memory and operating them as a single inference environment.
The Macs are connected using RDMA (Remote Direct Memory Access) over Thunderbolt 5, and Apple's machine learning framework MLX splits the model and runs it across them. RDMA is a mechanism that allows machines to access each other's memory directly and transfer data efficiently. Meta specifies macOS 26.2 or later as a requirement and calls for a configuration in which all the Macs used are directly connected to one another. Beyond securing the necessary number of machines and memory capacity, how the machines are connected is also an important condition for deployment.
As examples of operation on Mac, Meta presents application development with Muse Code and video-editing experiments. With the Mastodon client "Ice Cubes," the model carried out work from fixing bugs to implementing new features and building the app. In the video-editing software "DaVinci Resolve," it was tested on reviewing footage frame by frame, choosing highlights, and cropping them into vertical videos. Both are examples using real applications, but they do not reveal task success rates across a variety of projects or processing speed compared with cloud environments.
Options Widen Through the Apple Collaboration, but Challenges Remain for Self-Hosting
In an interview with Axios, Meta's AI chief Alexandr Wang said the company had worked with Apple for several months. Daniel Gross, a former Apple executive who now works at Meta, reportedly played an important role in building the relationship between the two companies. However, what has come to light is technical cooperation on running Muse Spark Open on Mac, not an announcement that it will be adopted in Apple Intelligence.
One of the biggest advantages of operating a model in your own environment is that input data can be processed without being sent to an external inference API. When handling sensitive code or internal documents, being able to control where data is stored and processed is significant. However, if an AI agent uses web search or external services, data sent to each of those services must also be considered. Running the model locally does not mean that no data from the overall system is sent externally.
Developers who cannot prepare large-scale hardware can still use the model through an API, as before. Meta points to availability through multiple inference service providers, and Baseten offers the NVFP4 version through Model APIs and dedicated inference environments. Its API uses an OpenAI-compatible interface and, the company says, supports tool calling, output of reasoning content, and contexts of up to 1 million tokens. Self-hosting and API use differ greatly in who takes on the procurement, maintenance, and operational management of the hardware that runs the model.
The license also needs to be checked. Meta says that, under the "Muse Model License," many developers can use the model free of charge under certain commercial-use conditions. But the release of the trained weights does not mean there are no restrictions at all on commercial use. Meta also says it applied additional safety training to the model, but the company's judgment in releasing it is a separate matter from the safety verification that users carry out for their own use cases.
The release of Muse Spark Open has widened the path from merely using a large AI model through an external API to running it directly on your own machines. However, given the scale of 816 billion parameters, self-hosting requires substantial capital investment and technical preparation. A key factor for future decisions will be how much context length and how many concurrent executions can be sustained with actual codebases and work histories. By measuring processing speed and operational burden under such conditions, organizations can judge which of the options—connecting multiple Mac Studios, using NVIDIA GPUs, or relying on an external API—best suits their needs.