Muse Spark Open

Frontier-scale open weights, downloadable today — tuned for agentic coding and tool use.

Visit website

Meta's 816B-parameter open-weight MoE model: 42B active per token, 1M-token context, downloadable since October 15, 2026.

Muse Spark Open Hugging Face model page

Muse Spark Open Hugging Face model page

1 / 2

Overview

What is Muse Spark Open?

Frontier-scale open weights, downloadable today — tuned for agentic coding and tool use.

Muse Spark Open is Meta's open-weight release of the Muse Spark model family, announced October 15, 2026. It is a mixture-of-experts model with 816 billion total parameters (42 billion active per token), a 1-million-token context window, and native multimodal input: text, image, audio, and video in, text out. Meta describes it as the most capable US open-weight model to date, a claim based on the company's own evaluation table.

The weights were downloadable from day one in BF16 and NVFP4 formats across Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Meta demonstrated the model running on two M5 Ultra Mac Studios with 512GB of unified memory each through Apple's MLX framework; the model card lists at least 700GB of accelerator memory for the NVFP4 build. Day-one serving support covers vLLM, SGLang, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth, and seven hosted API providers serve it, including Together AI, Fireworks AI, Groq, Baseten, Modal, Nebius, and OpenRouter.

The model ships under the Muse Model License: free for most commercial use including fine-tuning, distillation, and redistribution, with products required to carry a "Built with Muse Spark Open" attribution. Businesses whose model-serving revenue exceeds $20 million in any consecutive 12 months must sign a separate agreement with Meta; internal deployment and research are exempt.

Platforms and languages

Muse Spark Open Availability

Platforms

WebApi

Languages

English

Capabilities

Muse Spark Open Key Features

816B MoE, 42B active per token

Mixture-of-experts design with 816 billion total parameters; 42B run per token, with a 1M-token context window.

Natively multimodal input

Accepts text, image, audio, and video inputs and generates text output.

Downloadable weights since day one

BF16 and NVFP4 checkpoints on Hugging Face, SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

Broad serving support

Day-one support for vLLM, SGLang, TensorRT-LLM, MLX, llama.cpp, and Unsloth; hosted APIs via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

Desktop-scale deployment demo

Meta demonstrated the model running on two M5 Ultra Mac Studios with 512GB unified memory each via Apple's MLX framework.

Muse Model License

Free for most commercial use including fine-tuning and redistribution; model-serving businesses over $20M annual revenue need a separate agreement.

Best for

Who uses Muse Spark Open?

Developers and ML teams

self-host or fine-tune a frontier-scale open-weight model on their own infrastructure

Enterprises with procurement constraints

deploy a US-developed open model for agentic and tool-calling workloads

Plans and access

Muse Spark Open Pricing

Free

Common questions

Muse Spark Open FAQs

Is Muse Spark Open really free to use?

The weights are free to download, and the Muse Model License permits commercial use, fine-tuning, and redistribution for most users. Businesses earning over $20 million a year from model-serving need a separate agreement with Meta.

Can I run it on my own computer?

Meta demonstrated it on two M5 Ultra Mac Studios with 512GB of unified memory each. The model card asks for at least 700GB of accelerator memory for the NVFP4 build, so in practice this is multi-GPU server or high-end workstation hardware.

How is it different from the closed Muse Spark model?

Muse Spark Open is the open-weight release of the Muse Spark family with downloadable checkpoints, while Meta's hosted Muse Spark and Muse Wonder models remain API-only products.

Where can I try it without downloading the weights?

Seven hosted providers serve it via API, including Together AI, Fireworks AI, Groq, Baseten, Modal, Nebius, and OpenRouter — OpenRouter lists it at $1.25 per million input tokens and $4.25 per million output tokens.

Reviews

See what the community thinks and share your experience.

—

Based on 0 ratings

Rating distribution

0
0
0
0
0

Leave a review

Sign in to rate

Community reviews

No reviews

No written reviews yet.

Explore by category

AI Creation Generation

View all AI Creation Generation websites