AI
Meta releases Muse Spark Open, an 816B open-weight model
A giant model with downloadable weights, per its Hugging Face card.
Meta's Hugging Face model card (Oct 15) lists Muse Spark Open at 816B total parameters (the numbers a model learns) with 42B active at a time, a 1M-token context at inference (how much text it can read at once), and weights in BF16 and NVFP4, two ways of storing those numbers. The card says NVFP4 needs "at least 700GB of VRAM" (graphics-card memory) and BF16 about 2TB.
The license is the Muse Model License, which requires displaying "Built with Muse Spark Open." The card lists scores of 89.2 on MCP-Atlas, 80.4 on OSWorld Verified and 76.4 on Terminal-Bench 2.1. The caveat: it is nine days old, and we read the card but ran nothing, so treat the scores as the listing's numbers.
On your desk: not runnable at home, but an open-weight option on third-party hosts. If an n8n workflow ever swaps in a hosted model, the credit-line license term is the first thing to check.
Read the original at Hugging Face: Muse Spark Open 816B A42B model card →