Meta Releases 816B Parameter Muse Spark Open

Meta Superintelligence Labs has published Muse Spark Open, a mixture-of-experts model featuring 816 billion total parameters with 42 billion active per token. Built with a 1-million-token context window, the natively multimodal system handles text, image, video, and audio inputs. The unquantized BF16 weights weigh roughly 822 GB, alongside a compressed 484 GB NVFP4 variant distributed on Hugging Face under the commercial-use Muse Model License.

In a notable hardware demonstration, Meta verified that the model can run locally across two Apple M5 Ultra Mac Studio machines using Apple MLX, drawing on their combined 1 TB of unified memory and up to 1.2 TB/s memory bandwidth per system. Hosting is supported via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, alongside Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Serving environments span SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth, and PyTorch using TorchTitan. To optimize inference, Meta bundled auto-tunable Helion GPU kernels, yielding 1.03× to 3.95× speedups at the kernel level and a 3% end-to-end gain on AMD MI350 hardware.

For developers building long-running software engineering agents and multi-step reasoning systems, this release offers an open alternative to proprietary, closed-source models. Meta demonstrated these capabilities by using Muse Code to modify the Ice Cubes open-source iOS Mastodon app and control DaVinci Resolve for video editing. Although external validation on benchmarks such as SWE-bench and Terminal-Bench remains forthcoming, engineers can evaluate performance across full agent trajectories, time to first token, and prompt-injection resilience.