Meta released the weights for Muse Spark Open on Oct. 15, calling it the most capable U.S. open-weight model to date. Open weights mean anyone can download and run the trained model on their own hardware.
It’s a mixture-of-experts model with 816 billion total parameters, 42 billion of them active for each token, and a context window of 1 million tokens, according to Meta’s research blog. Mixture-of-experts models only use part of their network at each step, which cuts computing costs.
Meta says the model is built for coding and agent work and accepts more than text. The Batch, DeepLearning.AI’s newsletter, lists image, video and audio input.
Where it runs
Muse Spark Open is available through API providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. OpenRouter lists the price at $1.25 per million input tokens and $4.25 per million output tokens, The Batch reported.
Meta says it’s the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with Apple’s MLX framework. Each machine’s GPU can use up to 512 GB of memory, so two of them can hold the whole model, and Meta says the pair behave “like one supercomputer with a terabyte of unified memory.” The full-precision checkpoint is about 1.5 TB, according to The Signal, a newsletter by Shashwat Kumar.
Meta says it tested agent tasks entirely on that setup with Muse Code, its coding agent. One test shipped bug fixes and new features to Ice Cubes, an open-source Mastodon client for iOS. Another drove the DaVinci Resolve video editor, with the model reviewing footage frame by frame and cropping the best moments to vertical video.
How it compares
Artificial Analysis, an independent testing firm, gave the model 39 on its Intelligence Index, supporting Meta’s claim that it leads U.S. open models, The Batch reported. In Meta’s own comparison table, it leads on MCP-Atlas, a tool-use test, with 89.2, and on JobBench with 61.8.
It trails the best Chinese open models on some coding tests. It scores 76.4 on Terminal-Bench 2.1, behind Kimi K3 at 85.0, GLM 5.3 at 83.9 and Qwen3.8 at 82.0, according to The Batch.
Meta says the model had extra safety training because open weights carry their own risks, and that it met the release thresholds in Meta’s Superintelligence Scaling Framework. A full evaluation report is coming, the company says.
The license has conditions. Meta calls the Muse Model License “limited commercial terms” that let many developers use the model for free. The Signal reports that it allows use, changes, fine-tuning, distillation and resale, but every product built on it must display “Built with Muse Spark Open.”
Companies that offer the model to others as a service, with group revenue above $20 million in any 12-month period, must notify Meta and sign a separate agreement first, The Signal reported. Internal use is exempt. The attached usage policy also bars the unlicensed practice of professions including financial, legal and medical work, the newsletter said.
Source: Meta AI Research, Meta, “Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date”
Additional sources: The Batch (DeepLearning.AI), DeepLearning.AI, “Data Points: Muse Spark Open leads U.S. open models”; The Signal, Shashwat Kumar, “Open weights now come with fine print. Read it before the benchmark”
