Meta Releases 816B Parameter Muse Spark Open Model

Meta has officially released Muse Spark Open, an 816B total and 42B active parameter open-weights model featuring a 1M token context window. Built for natively multimodal coding and agentic operations, the release positions Meta as the leading American provider in the open-weight ecosystem. The deployment arrives alongside Meta's proprietary frontier offerings, Muse Wonder and Muse Wonder Full, as well as its 30B agentic model, Muse Glimmer.

The performance of Muse Spark Open places it above GLM 5.2 and near DeepSeek V4.1 Flash on public evaluations, though still behind leading Chinese models like GLM-5.3, released on Aug. 18th, and Kimi K3, released on Jul. 16th. Observers noted that safety post-training—likely a tuning of Muse Spark 1.2 rather than Muse Spark 1.3—resulted in a minor 1-point drop on the Artificial Analysis Intelligence Index to harden default safety behaviors. Meanwhile, competitors are pushing performance benchmarks, with Mistral Large 4 claiming an index score of 38 despite unreleased weights, Xiaomi V2.6 Pro leading overall open weights, and closed systems like Anthropic's Opus 5.5, OpenAI's GPT-6-Astra, and Google's Gemini-4-Argon defining the proprietary frontier.

For developers and enterprise practitioners, Muse Spark Open presents both capabilities and operational restrictions. While performance is capable enough for immediate adoption into agentic workflows, Meta's custom license imposes strict requirements: commercial products built on the model must display a "Built with Muse" label, and organizations generating over $20M per month in revenue must negotiate a separate commercial license. Additionally, the license includes acceptable use policies and restrictive terms on outputs and derivatives, creating legal considerations around synthetic data generation even as internal research usage remains permitted.