Meta’s Muse Spark Open has pulled American open-weight AI back toward the performance frontier after a period in which Chinese releases dominated the category. Interconnects’ analysis says the gap between the strongest US and Chinese open models has roughly halved, making American options relevant again for users who need high performance.
The 816-billion-parameter mixture-of-experts model activates 42 billion parameters for each token and handles text, images, video and audio as input. Its full BF16 weights occupy about 1.6 terabytes, while Meta’s compressed checkpoint is roughly 530 gigabytes. That still requires substantial hardware: Meta says the compressed version can fit on four Blackwell Ultra GPUs, eight H200s or two 512 GB Mac Studios.
Benchmark results are mixed rather than uniformly dominant. The model leads some open-model rankings and performs strongly on knowledge tests, but it trails peers on several agent evaluations. Interconnects also notes that Western open models often attract more adoption than similarly capable Chinese models.
The release is not unrestricted. Meta’s custom license requires attribution, and model-hosting businesses above $20 million in annual revenue need a separate agreement. The result is a stronger US open-weight option, but one with meaningful infrastructure and licensing constraints.