Meta Releases Open Weights for Muse Spark AI Model
Meta Platforms Inc. has officially launched the open weights for its Muse Spark artificial intelligence model, which had remained proprietary since its introduction in April. The new release, named Muse Spark Open, is specifically designed for coding and long-running agent tasks, positioning itself as the most capable open-weight model developed in a U.S. lab to date.
This model boasts a unique architecture wherein only 42 billion of its 816 billion parameters are activated per token input. Users can submit prompts that combine text with images, audio, or video, with Meta providing a context window that can accommodate up to 1 million tokens.
Muse Spark Open builds on Muse Spark 1.2, which was released in August, and includes enhanced safety features post-training to ensure robustness without relying on Meta’s proprietary services. According to Meta, Muse Spark Open has successfully passed the release criteria outlined in its Meta Superintelligence Scaling Framework, with a comprehensive evaluation report anticipated soon. A technical write-up from Hugging Face Inc. notes that training was informed by realistic developer workflows involving multi-file refactoring and extended debugging sessions.
In terms of performance testing, Meta evaluated the model outside of its data center environment, utilizing two Apple Inc. Mac Studio desktops equipped with M5 Ultra chips, which were interconnected via Apple’s MLX framework and provided nearly one terabyte of memory. During testing, the Muse Code functionality was demonstrated as it facilitated code modification and feature implementation for the open-source Mastodon client, Ice Cubes, while also analyzing video footage in DaVinci Resolve to efficiently select and crop ideal shots for vertical video content.
In Meta’s benchmark assessments, Muse Spark Open achieved a score of 61.8 on the JobBench agent test, significantly outperforming the 31.1 score of Inkling, an open model from Thinking Machines Lab Inc. Nvidia Corp.’s Nemotron 3 Ultra lagged behind with a score of 27.2. However, when it comes to coding, Chinese open models outperformed Muse Spark Open, with Kimi K3 leading the Terminal-Bench 2.1 coding assessment at 85, and GLM 5.3 and Qwen3.8 scoring higher than Muse Spark Open’s 76.4.
Conversely, on tool utilization, Muse Spark Open excelled with a score of 89.2 on the MCP-Atlas assessment, surpassing all competitors listed by Meta, including the Chinese models. This benchmark evaluates the model’s ability to complete plain-language tasks without specifying the tools required, leaving the selection of appropriate tools from 220 options across 36 active Model Context Protocol servers to the model itself.
Alongside the model weights, Meta has introduced a new licensing framework called the Muse Model License, characterized as carrying “limited commercial terms.” Developers interested in modifying the model or integrating it into their software will not incur any fees; however, a separate agreement is required for companies that intend to offer the model as a hosted service and exceed $20 million in gross revenue within a 12-month period. Additionally, products developed using the model must carry a label stating “Built with Muse Spark Open,” while military and critical infrastructure applications are explicitly prohibited.
Previously, Meta had released the much smaller Muse Glimmer model, which contained 30 billion parameters, in August. The company’s top-tier systems remain proprietary, with models like Muse Wonder only accessible through the Meta Model API and Muse Code. At launch, seven inference providers, including Groq Inc. and Fireworks AI Inc., will offer hosted access to the Muse Spark Open model. Developers can also download the weights from Hugging Face as well as the model catalogs of Amazon Web Services Inc., Google LLC, and Microsoft Corp.