Meta releases Muse Spark Open: an 816B-parameter multimodal model built for coding agents

Meta has released Muse Spark Open, a downloadable multimodal model aimed at coding, tool use and long-running agent workflows. The mixture-of-experts system contains 816 billion parameters but activates about 42 billion for each token. It accepts text, images, video and audio, produces text, and supports a context window of up to one million tokens.

Open weights do not mean a small local model

An open-weight release lets developers download and modify the trained parameters, although it does not necessarily reveal the training data or make the software unrestricted. The official model card lists the Muse Model License rather than a conventional open-source software license and sets conditions on use. Teams should review those terms before treating the checkpoint as commercially interchangeable with permissively licensed code.

Muse Spark Open uses a mixture-of-experts architecture: a router selects eight of 256 experts for each token. That reduces active computation compared with using every parameter at once, but all weights still have to be stored. Meta offers BF16 and compressed NVFP4 checkpoints. The model’s scale explains why the phrase “run locally” needs qualification.

Meta’s deployment guidance says the NVFP4 version can run across two Mac Studios, each equipped with an M5 Ultra and 512 GB of unified memory, or four 256 GB systems connected through Thunderbolt 5 and RDMA. It also supports server frameworks including SGLang, vLLM, TensorRT-LLM, Transformers and llama.cpp across several hardware vendors. For most individual developers, a hosted inference provider will be more practical than buying a one-terabyte desktop cluster.

What the model is designed to do

Meta positions Muse Spark Open for reading repositories, writing and debugging code, invoking tools and recovering from failures over extended tasks. Its perception encoder lets the same model inspect screenshots, documents, video and audio while producing a textual response. A separate commentary channel can state intended actions before tool calls, which can help an orchestration layer expose what the agent plans to do.

The release is derived from the closed Muse Spark model introduced in August. Meta says it added safety post-training for standalone use, while the model card still recommends deploying it inside a wider system with guardrails matched to the application. That warning is important: downloadable weights remove a service provider’s runtime controls, so the organization operating the model assumes more responsibility for permissions, monitoring and content policy.

Benchmarks are useful, but not the whole comparison

The Guardian’s launch report says Meta claims the model rivals or exceeds leading Chinese open-weight alternatives. Meta’s evaluation methodology combines official leaderboards and external evaluations with internal runs and some competitors’ self-reported scores. The source and settings therefore need to be checked benchmark by benchmark. They do not establish cost, latency or reliability for a buyer’s own repository and toolchain.

The practical development is therefore access: an advanced US-built model with downloadable weights, broad input modalities and documented serving options is now available. Its openness improves inspection and customization, but the custom license, very large memory footprint and need for application-level safeguards remain central constraints. A serious evaluation should reproduce the relevant tasks, record tool failures and compare total serving cost—not choose the model from a headline benchmark alone.