Meta released Muse Spark Open on Hugging Face, an 816B total parameter model with 42B active parameters at inference, using a Mixture-of-Experts Transformer with a Perception Encoder. The release includes BF16 and NVFP4 checkpoints and supports text, image, video, and audio inputs with a 1 million token context window, targeting coding and agentic workloads. Apple optimized the model for the MLX framework; running the full model requires two linked M5 Ultra Mac Studios. The 4-bit NVFP4 quantization requires at least 700GB of GPU memory, while a 3-bit version fit on a 512GB M3 Ultra at roughly 344GB and ran at about 21 tokens per second on short prompts.
The open release shares the same base as the proprietary Muse Spark 1.3 but includes additional safety post-training to improve robustness when deployed without system guardrails, reversing the closed-weight strategy Meta announced in April. The model is also available through Azure AI Foundry and hosted inference providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Self-hosting tools include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, and llama.cpp. Meta's Helion GPU kernels deliver kernel-level speedups of 1.03 to 3.95 times, though the end-to-end improvement on AMD MI350 hardware is 3%.
The model is released under the custom Muse Model License, which is not OSI-approved. All licensees must display a visible 'Built with Muse Spark Open' notice. Only model-as-a-service providers whose revenue exceeds $20 million in any twelve-month period must obtain a separate agreement with Meta; internal use, academic research, and products embedding the model as a feature are exempt. Hugging Face access requires submitting full legal name, date of birth, and full organization name with corporate identifiers.
Meta claims Muse Spark Open is the most capable US-developed open-weight model, a narrower claim than global leadership. On Artificial Analysis benchmarks, Muse Spark Open scores 39, while MiMo-V2.6-Pro and Qwen-3.8 sit at 46. Interconnects.ai placed its performance between GLM 5.2 and GLM 5.3/Kimi K3, and the model ranks eighth on the Interconnects Adoption Dashboard. Interconnects.ai argued that Western open-weight models historically receive substantially more adoption than Chinese models with equivalent performance, giving Muse Spark Open a structural advantage, while also arguing that complex non-standard licenses are an unnecessary strategic risk for labs trying to regain adoption. The Zvi newsletter noted the model can be used in contexts where running Chinese-origin models is not acceptable.
For the closed Muse Spark 1.0, Artificial Analysis placed it fourth on their Intelligence Index behind Gemini 3.1 Pro Preview, GPT-5.4, and Claude Opus 4.6. It scored 39.9% on HLE, trailing Gemini 3.1 Pro Preview (44.7%) and GPT-5.4 (41.6%), and ranks fifth on CritPT at 11%. On the GDPval-AA agentic benchmark it scored 1427, below Claude Sonnet 4.6 (1648) and GPT-5.4 (1676). DataCamp noted weaknesses on ARC-AGI-2 and Terminal-Bench 2.0. Meta collaborated with more than 1,000 physicians to curate health-focused training data; Muse Spark scored 42.8 on HealthBench Hard, outperforming GPT-5.4 (40.1) and Gemini 3.1 Pro (20.6). DataCamp noted Muse Spark used 58 million output tokens on Artificial Analysis's run versus 157 million for Claude Opus 4.6.
A newer closed-weight version, Muse Spark 1.1, scores 88.4 on CharXiv Reasoning with tools, below Claude Opus 4.8's 89.9 and Muse Spark 1.0's 88.9. On BabyVision with tools it scores 76.3, trailing GPT-5.5 at 83.6 and Claude Opus 4.8 at 81.2. Muse Spark 1.1 leads the health category with 59.3 on HealthBench Professional, ahead of Claude Opus 4.8's 55.8. On safety, it achieves 97.7 on BioTIER refusals and 99.8 on Chemical Agents refusals, and matches GPT-5.5 on adversarial robustness with a 0.5 attack success rate on StrongREJECT v2.