# Muse Spark Open explained: what Meta has said so far

> Meta released Muse Spark Open, an 816B-parameter open-weight multimodal model, claiming it is the top US open-weight model.

*Meta has released the weights of a large multimodal model for coding and agent work, but its performance claims have not yet been independently tested.*

By Bez · SuggestedTech
Canonical: https://suggestedtech.com/news/muse-spark-open-explained-what-meta-has-said-so-far

Meta has released Muse Spark Open, an "open-weight" AI model. That means the trained model files, called weights, are public, so developers can download and run them. Meta calls it "the most capable US open-weight model to date", but that claim has not been independently checked yet.

## What happened

Meta AI Research announced the model on its [company blog](https://research.meta.ai/blog/muse-spark-open) on 15 October. The quote above comes from that post.

Here is what Meta says about the model:

- It is natively multimodal, meaning it was built from the start to handle more than one kind of input, not just text.
- It has 816B total parameters (the numbers a model learns during training). Only 42B are active at a time, because it uses a "mixture of experts" design that switches on only part of the model for each task.
- It has a 1M-token context window, which is how much text it can consider at once.
- It focuses on coding and agentic work, meaning long tasks, complex software engineering and multi-step reasoning.

Meta says the model received extra safety post-training, the stage where a model is tested and tuned before release. It also says the model met the release thresholds in its Meta Superintelligence Scaling Framework. A full evaluation report is "available soon", according to Meta.

The blog's benchmark section gave no figures in the text we could review. Until outside testers publish results, treat the performance claims as Meta's own.

## What it means for you

If you use AI tools but don't build them, you may meet this model through other products or services that choose to adopt it. The notes don't say whether any consumer product uses it.

If you are a developer, the cost picture is mixed. Weights are released under the Muse Model License, which has limited commercial terms. Meta says many developers can use the model free of charge. Read the licence before building a business on it.

You can also run it on your own hardware. Meta says two Apple M5 Ultra Mac Studios, linked together, can run it through MLX, Apple's machine-learning framework. Each machine has up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth, and together they act like one machine with about 1 TB of memory.

## How to try it

You have several routes, depending on how technical you are.

- **Download the weights** from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden or Microsoft Foundry Models.
- **Use a hosted API**, which lets you send requests without running the model yourself, from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter or Together AI.
- **Run it yourself** with SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp or Unsloth.

Meta also says Helion kernels in SGLang give 1.03–3.95x kernel speedups on Nvidia and AMD GPUs. Kernels are the small programmes that do the heavy calculations on a graphics chip.

## Key takeaways

- Meta released Muse Spark Open, a model whose weights anyone can download under a licence.
- Many developers can use it free, but the licence has limited commercial terms.
- Meta's claim of being the most capable US open model is untested by outsiders.

## Sources

- [Muse Spark Open](https://research.meta.ai/blog/muse-spark-open) — Meta AI Research, 2026-10-15