Meta Releases Muse Spark Open, the Most Capable US Open-Weight AI Model to Date
Meta has released Muse Spark Open, an 816-billion-parameter multimodal model with 42 billion active parameters and a 1-million-token context window, calling it the most capable US open-weight model to date. The model is available today through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with weights downloadable on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.
Desktop Deployment on Apple M5 Ultra Mac Studios
Muse Spark Open runs on desktop systems using two Apple M5 Ultra Mac Studios with the MLX framework, leveraging unified memory architecture that allows each Mac Studio’s integrated GPU to use up to 512 GB of memory. This setup provides a terabyte of unified memory and 1.2 TB/s memory bandwidth, enabling the two machines to function as a single supercomputer for on-device inference. The model is optimized for inference engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with specific optimizations for AMD, Apple, Intel, and NVIDIA silicon.
Performance in Coding and Agentic Workflows
Meta tested agentic workflows entirely on-device using Muse Code powered by Muse Spark Open, including shipping real iOS changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve end to end by reviewing footage frame by frame to identify and crop strong moments vertically. The open-weight version includes custom, auto-tunable Helion GPU kernels in SGLang, delivering 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350 hardware.
Safety Post-Training and Open-Weight Considerations
Muse Spark Open includes additional safety post-training for open-weight release considerations, with Meta stating the results demonstrate strong safeguards across measured risk categories. The company notes that releasing open weights is part of advancing safety work across the field while bringing personal superintelligence to broader access.
Comparison with Global Open-Weight Models
In Meta’s own comparison, GLM 5.3 (45), Kimi K3 (44) and Qwen3.8 (40) rank ahead of Muse Spark Open on the Artificial Analysis Intelligence Index. Muse Spark Open scored 39 points on the same index, compared to Mistral Large 4’s 38 points, placing the models neck and neck for the best open-weight model outside China. Muse Spark Open is not yet listed in Artificial Analysis’s ranking of open-weight models. If the 39 points are confirmed, it would tie DeepSeek V4.1 Flash for fifth place and become by far the best American model in the ranking. Meta’s model leads on agent and tool tasks such as JobBench and MCP-Atlas, according to internal evaluations.
Multimodal Capabilities and Use Cases
Muse Spark Open is a natively multimodal model, designed to understand and process images and video alongside text. This enables it to perform complex tasks that require reasoning across different data types, such as analyzing visual content while generating or modifying code. Meta demonstrated this capability by using the model to fix bugs in an iOS application autonomously and to review video footage frame by frame in DaVinci Resolve to identify and crop strong moments vertically.
The model’s multimodal design supports advanced agentic workflows where perception, reasoning, and action are integrated. For example, in software development, it can interpret UI screenshots, suggest code changes, and validate fixes—all within a single continuous process. In video editing, it can analyze temporal sequences, detect salient visual elements, and apply edits based on semantic understanding rather than just pixel-level patterns.
These capabilities are built into the model’s architecture and are accessible through standard inference engines like SGLang, vLLM, and Hugging Face Transformers. Developers can leverage the multimodal input handling to build agents that interact with graphical interfaces, process multimedia logs, or assist in creative workflows involving both code and media. The 1-million-token context window further allows the model to maintain coherence over long sequences of multimodal data, making it suitable for extended reasoning tasks.
Sources
- Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date
- ‘Doing What It Takes’ in AI. OpenAI, Meta & TSMC. ARD #185
- Meta’s Muse Spark Open Comes With Strings Attached | Trending Topics
Photo: LPS.1 (CC0)
