On October 15, 2026, Meta released the AI model Muse Spark Open. It distributes the weights (the trained data that makes up the model) of Muse Spark, which Meta has been using in its own services since April 2026, so that anyone can download it and run it on their own systems. Meta says the model is aimed at programming and at "agent" use, where AI works through tasks while using tools.
What is Muse Spark Open?
Muse Spark is the model Meta has been using in its services since April 2026. It is separate from Muse Wonder, the closed higher-end model announced on October 12. According to Meta, Muse Spark Open is based on Muse Spark 1.2, released in August, with additional training to improve safety. Meta says it shares the same base model as the current Muse Spark 1.3.
The main specifications are as follows.
- Architecture: MoE (Mixture of Experts). It has 816 billion (816B) parameters in total, of which 42 billion (42B) are used in each pass
- Amount it can read at once (context): 1 million tokens
- Input: text, images, audio and video
- Output: text
MoE is a design in which, even though the whole model is large, only some "expert" parts are activated for each pass, so it runs faster and more cheaply than its size would suggest.
Where to get it and how to use it
The weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud's Model Garden and Microsoft Foundry. On Hugging Face, you must enter your name, date of birth and affiliation before downloading.
Even without setting up your own servers, you can use it through APIs from providers such as Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI.
To run it on your own systems, you need more than 700GB of GPU memory in a lightweight format (NVFP4), and 2TB in the original format (BF16). Meta also explains that it can run with MLX, Apple's framework, by connecting two M5 Ultra Mac Studios with 512GB of memory each.
License terms
The license is Meta's own Muse Model License, which Meta describes as a license that attaches certain conditions to commercial use. Companies with total revenue of more than $20 million over the most recent 12 months must notify Meta and sign a separate agreement if they offer the model to other companies as a service, such as through an API. Products and services using the model must also prominently display "Built with Muse Spark Open," and the terms of use prohibit uses such as military applications. Because it differs from an open-source license that anyone can use without restrictions, you should check the terms before using it for business.
Performance evaluation
The main figures Meta listed in its model card are as follows. Muse Spark Open was measured with its reasoning depth set to the deepest setting (xhigh).
- Artificial Analysis Intelligence Index: 39. Among the open models used for comparison, GLM 5.3 scored 45, Kimi K3 44, Qwen3.8-2.4T 40 and Mistral Large 4 38
- OSWorld Verified (computer operation): 80.4
- MCP-Atlas (tool calling): 89.2
- GPQA Diamond (graduate-level science questions): 89.0
- Terminal-Bench 2.1 (command-line tasks): 76.4
- DeepSWE 1.1 (software development): 60
On some items it falls short of Chinese open models. For example, Kimi K3 scored 68 on DeepSWE 1.1 and 85.0 on Terminal-Bench 2.1. On the overall index, too, it trails GLM 5.3 and Kimi K3. On the other hand, it far outperforms the U.S. open models in the same table, Inkling (26 overall) and Nemotron 3 Ultra (23), and it posts the highest figures in the table on JobBench (61.8) and MCP-Atlas.
CoAI's take
Muse Spark Open is an example of Meta taking a model it has used in its own services, giving it additional safety training, and releasing it complete with its weights. It is a large model for an open model from a U.S. company, and since it is immediately available not only on Hugging Face but also from major clouds and API providers, there are plenty of options for using it.
That said, even by Meta's own figures, its overall performance does not reach that of Chinese open models such as GLM and Kimi. Running it requires a server with large GPUs, and the license comes with conditions. It is best to decide first whether you value being able to run it on your own systems or the convenience of an API, and then test and compare it on your own work.
Sources
- Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date (Meta)
- Muse-Spark-Open-816B-A42B (Hugging Face)
- Muse Spark Is Now an Open-Weights Model (PyTorch Blog)
- Meta releases 816B-parameter Muse Spark Open for coding and agent work (SiliconANGLE)
- Meta launches new Muse Spark Open model, optimized to run on Mac Studio (9to5Mac)
