Muse Spark Open
Meta · 2026-10-15
Meta's open-weights, natively multimodal mixture-of-experts model for coding and agentic work, released under the Muse Model License with limited commercial terms; it takes text, image, video and audio input with text output, serves a 1,048,576-token context in Meta's vLLM recipe, and supports low, medium, high and xhigh reasoning strength, with xhigh the default. The model card reports 816B total and approximately 42B active parameters (256 experts, 8 routed per token) plus a 1.8B-parameter perception encoder. The official config and safetensors index at revision 0fb02478b02565b2ce7fc57e6051c91172774ac9 are gated, and Hugging Face's raw 822,036,529,664-element tensor count (including F32 tensors) does not establish a logical total, so total_parameters is left undeclared.