Meta Unveils Muse Spark Open, a 816‑Billion‑Parameter Multimodal Model with Restricted Licensing

Meta's coding-optimized model, Muse Spark Open, is now accessible in Amazon SageMaker JumpStart, supporting a 1 million token context window and handling various input modalities. The model, trained on real developer workflows, can be deployed with ease and offers two quantization variants.

Meta introduces Muse Spark Open, a model with 816 billion parameters, capable of processing text, images, audio, and video. Designed for agentic coding, it can read codebases, write and debug code, and call tools, offering a comprehensive guide for developers.

Latent Space's AINews newsletter covers recent AI model releases including Meta's Muse Spark Open, an 816-billion parameter Mixture-of-Experts model with open weights, and Reflection Beam 1.0 released under Apache 2.0. The report details Muse Spark Open's architecture, performance benchmarks against competitors like Nemotron and GLM-5.3, licensing terms requiring attribution for commercial users, and availability across platforms including Apple MLX and vLLM. It also notes community reactions, including criticism regarding model availability and speculation about withheld versions.

Meta introduced Muse Spark Open, an open-weight model with 816 billion parameters and a 1 million token context window, designed to be natively multimodal and run locally on Apple computers. While it requires at least 1 terabyte of RAM, it performs best on a Mac Studio equipped with the M5 Ultra chip. Meta demonstrated its capabilities by integrating it with Muse Code for coding fixes on a Mastodon client and using it to review and crop footage in DaVinci Resolve. Open-weight models offer cost efficiency, transparency, and flexibility compared to closed proprietary models, and the model is now available via various API platforms.

Meta introduced Muse Spark Open, which it claims is the most capable US open-weight model, featuring 816 billion parameters (42 billion active), up to one million tokens of context, and support for text, image, video, and audio. Meta states that Muse Spark Open scores ahead of Qwen 3.8, GLM 5.3, and Kimi K3 on benchmarks like JobBench, MCP-Atlas, HLE-Diamond, and ZeroBench in Agentic, Knowledge & Reasoning, and Multimodal categories, though it falls short of paid Muse Wonder Full models and some competitors on coding benchmarks. The model is optimized for Apple's M5 Ultra and weights are available on Hugging Face under a free license, but providers earning over $20 million annually from API services require a separate agreement with Meta.