Meta releases Muse Spark Open, an 816B open-weight model for coding agents

What happened

On October 15, Meta Superintelligence Labs released the weights for Muse Spark Open, calling it the most capable US open-weight model to date. The weights are available for download under the Muse Model License.

Key facts

Size: 816B total parameters with 42B active, natively multimodal, with a 1M-token context window. Focus: Coding, long-horizon agent work, complex software engineering, and multi-step reasoning, per Meta's announcement.

Desktop run: Meta says it is the first model in its class that can run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX stack (unified memory up to 512 GB per Studio). Hosted APIs today: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

Weights: Download from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with AMD, Apple, Intel, and NVIDIA paths listed.

License: Muse Model License and model card with limited commercial terms; Meta says many developers can use it for free. Safety: Meta says it ran post-training for open release and that the model met its Superintelligence Scaling Framework release thresholds; details are on the model card, with a fuller evaluation report "available soon."

Vendor benchmark claims belong on Meta's evaluation methodology report; treat them as company-reported until you re-run them yourself.

How to use it

Open-weight frontier models let teams fine-tune, air-gap, and control cost in ways closed APIs cannot. An 816B MoE available through hosted providers or the desktop configuration Meta describes widens who can ship agentic coding tools without renting a closed frontier API for every token. Compare it side by side with Muse Wonder and other open options before you pick a stack.

Skim the official blog and the Hugging Face collection. For a quick try, hit one of the listed hosts (OpenRouter, Together, Fireworks, and the others above) rather than standing up hardware first. If you self-host, start from MLX on Apple silicon or SGLang/vLLM on NVIDIA/AMD, then tune with PyTorch / TorchTitan as Meta describes. Before you commit a product path, run Make a product feature comparison against Muse Wonder and your current coding model, and keep agent briefs tight with Write a task delegation brief. When Spark Open invents a bug in a PR, capture it with Write a reproducible bug report.