Morning Edition · Sunday, October 18, 2026Published at 2:13 AM EDT · New York

Meta Releases 816-Billion-Parameter Muse Spark Open, Which Trails Kimi K3 on Meta's Own Coding Table

The mixture-of-experts model activates 42 billion parameters per token and supports a 1-million-token context. DeepSeek's V4.1-Flash-1017 update, released two days later, costs $0.08 per million input tokens.

Meta published the weights of Muse Spark Open on October 15. It is a mixture-of-experts transformer with 816 billion total and 42 billion active parameters, a Perception Encoder for native multimodal input and a 1-million-token context wind…