Meta Releases Muse Spark Open, Claims Most Capable US Open-Weight Model
An 816B open-weight model released for free commercial use sets a high bar for accessible frontier AI capability. The Apple-Meta partnership enables local deployment on consumer Mac hardware for a frontier-class model that previously required cloud infrastructure.
Meta released Muse Spark Open on HuggingFace on October 15, 2026, describing it as 'the most capable US open-weight model to date.' The model uses a mixture-of-experts architecture with 816 billion total parameters across 256 experts and 62 layers, routing each token through 8 to operate at 42 billion parameter speed. It supports a 1 million token context window and handles text, image, video, and audio natively. The open release shares the same base as the proprietary Muse Spark 1.3, with additional safety post-training, reversing a closed-weight strategy Meta had announced in April. Meta also noted it had released a closed frontier model the week before making this open release.
Meta partnered with Apple, which optimized the model for its MLX framework to run on two linked M5 Ultra Mac Studios. Running the full model requires at least two of those machines with a combined 1 terabyte of memory, costing tens of thousands of dollars. The 4-bit NVFP4 quantization needs at least 700GB of GPU memory; a 3-bit quantized version tested on a 512GB M3 Ultra consumed approximately 344GB and ran at about 21 tokens per second on short prompts. The license permits free commercial use with a 'Built with Muse Spark Open' attribution label; API sellers exceeding $20 million in annual revenue must negotiate a separate deal with Meta.
Artificial Analysis found Muse Spark scored 39.9% on HLE, placing third behind Gemini 3.1 Pro Preview at 44.7% and GPT-5.4 at 41.6%, and fifth on CritPT at 11%. On the GDPval-AA agentic benchmark it scored 1427, below Claude Sonnet 4.6 at 1648 and GPT-5.4 at 1676 but ahead of Gemini 3.1 Pro Preview at 1320; Artificial Analysis said 'agentic performance does not stand out.' Muse Spark 1.1, evaluated separately, scored 88.4 on CharXiv Reasoning with tools, below Claude Opus 4.8's 89.9 and its predecessor Muse Spark 1.0's 88.9, and 76.3 on BabyVision with tools, trailing GPT-5.5 at 83.6 and Claude Opus 4.8 at 81.2. It led the health category with 59.3 on HealthBench Professional, ahead of Claude Opus 4.8's 55.8, and matched GPT-5.5 on adversarial robustness with a 0.5 attack success rate on StrongREJECT v2.
Coverage on October 16 filled in hardware requirements for the 816-billion-parameter open-weight model released October 15: running the full model requires two linked M5 Ultra Mac Studios with a combined 1 terabyte of memory, costing tens of thousands of dollars.
A 3-bit quantized version tested on a 512GB M3 Ultra consumed about 344GB and ran at roughly 21 tokens per second. One source noted Meta had released a closed frontier model the week before this open release.
Meta released Muse Spark Open on October 15, an 816 billion parameter open-weight model on HuggingFace, calling it 'the most capable US open-weight model to date' and reversing a closed-weight strategy announced in April.
It shares its base with the proprietary Muse Spark 1.3; the license allows free commercial use with attribution, though API resellers over $20 million in annual revenue must negotiate separately. Artificial Analysis placed it third on HLE reasoning at 39.9%, behind Gemini 3.1 Pro Preview and GPT-5.4, and said agentic performance 'does not stand out,' scoring 1427 on GDPval-AA against GPT-5.4 at 1676 and Claude Sonnet 4.6 at 1648.
SensorTower data showed Muse app Day 1 retention beat ChatGPT by 32 points; 14-day retention trailed by 4.3 points
Muse Spark 1.1 benchmark analysis published, showing strong health and safety scores but trailing top competitors on some reasoning and vision tasks
Artificial Analysis published Muse Spark benchmark evaluation, finding strong reasoning scores but weaker agentic performance
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free