Muse Spark Open is now available, and its weights are being released as the most capable US open-weight model to date. It is natively multimodal, with 816B total parameters (42B active) and a 1M-token context window. Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
For the first time, an advanced model of this class can run on a desktop. Using two Apple M5 Ultra Mac Studios and the latest MLX framework, the machines pool their unified memory to hold the entire model. Apple's unified memory architecture lets each M5 Ultra's integrated GPU use up to 512 GB, and the chip's 1.2 TB/s of memory bandwidth reads the small slice each token needs. Testers ran agentic workflows entirely on-device, including shipping real iOS changes to Ice Cubes, an open-source Mastodon client, and driving Da Vinci Resolve end to end.
The model is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, Open Router and Together AI. Weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. It runs on engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon. Custom Helion GPU kernels in SGLang deliver 1.03-3.95× kernel-level speedups on NVIDIA and AMD GPUs.
The weights ship under the Muse Model License, which has limited commercial terms and lets many developers use the model for free. Additional safety post-training was applied, and the model met the release thresholds of the Meta Superintelligence Scaling Framework. A full evaluation report is expected soon.
Source: Hacker News · Summarized by HeadlinesBriefing