Meta Releases 816B-Parameter Muse Spark Open for Coding and Agent Work
Meta opens weights for its largest multimodal model, featuring 1M context and new commercial licensing terms.
ImportanceStructuralEvidenceE2 unreplicatedWrite-upQuick
On October 15, 2026, Meta Platforms Inc. released the open weights for Muse Spark Open, a multimodal large language model with 816 billion total parameters and 42 billion active parameters per inference. Designed specifically for coding and long-running agent tasks, the model supports mixed text, image, audio, and video inputs with a context window of up to 1 million tokens.
According to Meta's internal benchmarking, Muse Spark Open scored 61.8 on the JobBench agent test, surpassing Thinking Machines Lab's Inkling (31.1) and Nvidia's Nemotron 3 Ultra (27.2). In tool usage, it achieved 89.2 on MCP-Atlas, claiming superiority over all listed rivals. However, in pure code generation, it trails several Chinese open-source models: Kimi K3 scored 85 on Terminal-Bench 2.1, while GLM 5.3 and Qwen3.8 both exceeded Muse Spark Open's score of 76.4.
The accompanying Muse Model License allows developers to use the model for free when adapting it or integrating it into their own software. Companies selling the model as a hosted service must sign a separate agreement if their gross revenue exceeds $20 million over a 12-month period. Products built on the model must display a "Built with Muse Spark Open" label, and military or critical infrastructure uses are prohibited. Seven inference providers, including Groq and Fireworks AI, offer hosted access at launch, and weights are available via Hugging Face, AWS, Google Cloud, and Azure catalogs.