Meta releases Muse Spark Open, an 816B-parameter open-weight AI for code and media

Meta has released Muse Spark Open, an open-weight frontier AI model designed to process text, images, audio, and video within a single, massive architecture. This release marks a significant step in making powerful, multimodal AI systems accessible for developers and researchers to run and build upon, shifting the landscape of open-weight models.

What is Muse Spark Open?

Muse Spark Open is a Mixture-of-Experts (MoE) model with approximately 816 billion total parameters, of which about 42 billion are active for processing any given token, according to the announcement on the Hugging Face blog. It is a multimodal model, accepting text, images, audio, and video as input and generating text as output. Meta describes it as a "coding-optimized model" trained on real developer workflows like multi-file refactoring and debugging. A key feature is its support for a context window of up to one million tokens, which includes all input modalities.

How is the model designed and what are its capabilities?

The model's architecture is built around a language model core with 62 layers. Its Mixture-of-Experts design uses 256 experts per layer, routing each token through 8 of them, which is how it achieves high total parameter count with efficient active computation. Separate encoders handle different modalities: a perception encoder processes images and video frames, while a new audio encoder converts sound into tokens. The model is explicitly designed for "agentic" behavior, meaning it can plan, act, observe, and self-correct over long tasks, including browsing the web with visual understanding.

How can developers run and use it?

The model checkpoints are available on Hugging Face in two formats: a BF16 version (about 1.5 TB) and a 4-bit quantized NVFP4 version optimized for NVIDIA's Blackwell GPUs. According to the source, the Hugging Face transformers library now supports the model through MuseSparkForConditionalGeneration and MuseSparkProcessor. Running the full model requires significant hardware; the BF16 checkpoint is designed to run across 4 nodes of 8 H200 GPuses. The blog post provides code examples for text-only, image-and-text, and video inference, detailing how to use the chat template and control the model's "reasoning effort" from low to extra-high.

What does this release mean for AI development?

Muse Spark Open represents the current pinnacle of openly released model weights, described as "the most capable US open-weights model to date." Its scale and multimodal, agentic focus provide a new base for research and commercial development in complex AI assistants. By being open-weight, it allows the community to inspect, fine-tune, and deploy the model without API restrictions, though the computational requirements place it firmly in the domain of well-resourced organizations or cloud services. It also continues the trend of major labs releasing powerful models as open weights, fostering ecosystem growth around their architectures.

How does this relate to work in Grafy?

For developers and creators in the Grafy community, the capabilities of Muse Spark Open align closely with complex, multimodal projects. Its strength in agentic coding and long-context understanding is directly applicable within Grafy Code, our AI coding studio. Developers could leverage a locally or cloud-hosted instance of Muse Spark Open to power advanced code generation, refactoring, and debugging sessions over entire codebases within the Grafy environment. Furthermore, its multimodal reasoning could enhance research workflows in Grafy Reader, allowing for deep analysis across PDFs, images, and audio files in a single session.

What are the practical considerations for using it?

The primary barrier is computational. Hosting the 816B parameter model requires substantial GPU memory, though the active parameter count of 42B makes inference faster than a dense model of the same total size. The 1-million-token context window also demands a large Key-Value cache, estimated at about 266 GB for a full context in BF16 precision. The Hugging Face blog outlines deployment options using vLLM, llama.cpp, and SGLang to help manage these resources. For most users, accessing Muse Spark Open will likely occur through cloud inference endpoints or services that host the model, rather than local deployment.

Frequently asked questions

How is Muse Spark Open different from Muse Glimmer?

Muse Glimmer is a smaller, dense model from the same family, ideal for local use. Muse Spark Open is vastly larger, uses a Mixture-of-Experts architecture, and includes an audio encoder. They share a tokenizer and chat template format, but Spark Open is designed for frontier-scale tasks.

Can I run Muse Spark Open on a single consumer GPU?

No. The model requires hundreds of gigabytes of GPU memory just to load. The NVFP4 quantized version is designed for a single node of NVIDIA's flagship B200 GPUs, far beyond consumer hardware. Practical use requires multiple high-end data center GPUs or a cloud service.

What is the "reasoning effort" setting?

It's a control that adjusts how many internal reasoning steps the model takes before producing an answer. Settings range from "low" (16 steps) to "xhigh" (512 steps). Higher effort yields more thorough reasoning but consumes more tokens and time. The model cannot turn reasoning off entirely.

---

Source: Hugging Face Blog — Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta