In brief
- Meta released the weights of Muse Spark Open, a multimodal model with 816B parameters.
- Free download under the Muse Model License, with limited commercial terms.
- Self-hosting needs large GPU infrastructure. Without it, use an API provider from Meta’s list.
Meta released the weights for Muse Spark Open, a multimodal model with 816B total parameters (42B active) and a 1M-token context window.
AI Desk · Published October 16, 2026
First look: this report is based on Meta’s official announcement and documentation and has not yet been tested hands-on by Maitz.
Meta announced the release on October 15, 2026, on its AI Research blog. The company calls it the most capable open-weight model to date.
What did Meta announce with Muse Spark Open?
The model is built for coding and agentic work. Meta points to long-horizon tasks, complex software engineering and multi-step reasoning. It accepts text, image, video and audio, and outputs text.
Weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. API access runs through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
| Detail | According to Meta |
|---|---|
| Architecture | Mixture-of-Experts with a perception encoder |
| Parameters | 816B total, 42B active |
| Context window | 1M tokens |
| Input / output | Text, image, video and audio / text |
| Checkpoints | BF16 (about 1.65 TB) and NVFP4 (about 537 GB) |
| License | Muse Model License: use, modify, redistribute, with “Built with Muse Spark Open” notice |
| Release date | October 15, 2026 |
How much hardware does it need?
A lot. The model card says the NVFP4 version needs a GPU setup with at least 700 GB of VRAM, and BF16 needs 2 TB. The docs suggest a server with four or eight B300 accelerators.
Meta also stresses a desktop route: two Mac Studio machines with M5 Ultra chips, each with up to 512 GB of unified memory, running on Apple’s MLX framework. The company says this is the first model in its class to run that way.
Why does an open-weight release matter?
Open weights mean you can download, tune and run the model on your own infrastructure. Prompts and responses stay inside your systems. That matters most for teams with strict privacy rules.
There is a catch. The license allows use, modification and redistribution with two main conditions. Products must show “Built with Muse Spark Open”, and anyone offering models as a service through an API, with combined revenue above $20 million in a 12-month period, needs a separate agreement with Meta. According to Meta, this lets many developers use the model for free.
Meta also notes safety work. The open version received extra safety post-training, and the company says it met the release thresholds of its Meta Superintelligence Scaling Framework.
What we don’t know yet
- The full evaluation report. The model card includes a benchmark table, and Meta says the complete report will be available soon.
- Hosted API prices at each provider, which the announcement does not list.
What it means for you
- Developers: with GPU servers available, download the weights from Hugging Face and test. Without them, a hosted provider from Meta’s list is the simplest route.
- Companies: review the Muse Model License and Meta’s usage policy before any commercial use. Self-hosting keeps data in your own environment.
- Everyday users: nothing to install today. This is infrastructure, not an app. For Meta’s previous model, see our story on Muse Wonder.
- More AI tools: browse the Maitz AI tools directory.
More on Maitz
Sources
- Meta AI Research: Introducing Muse Spark Open, research.meta.ai, October 15, 2026
- Hugging Face: Muse-Spark-Open-816B-A42B Model Card, meta-models, October 15, 2026
- Meta Model API Docs: Muse Spark Open, Get the model, Run inference, dev.meta.ai
- Verified: October 16, 2026
