Reportedly: Meta releases Muse Spark Open, an 816B-parameter open-weights multimodal model
Reported by Zhidongxi · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Meta released Muse Spark Open, an open-weights model with 816B total parameters, 42B active per token, and 1M-token context. It supports multimodal input and targets coding and agent tasks. The model is available via API platforms and downloadable from Hugging Face, Zhidongxi reported.
China context
- Outside China
- Open weights · zhidx.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the weights from Hugging Face and run the model locally on Apple hardware, or use it via multiple API providers. Benchmark scores suggest it may underperform several Chinese open-weights models on general intelligence, so evaluate on your specific coding or agent tasks.
- For investors
- Meta's entry into open-weights multimodal models intensifies competition with Chinese labs, but its lower benchmark score may limit adoption unless it excels in specialized coding or agent benchmarks. Watch for enterprise adoption and API usage data.
Translated from Chinese. Quotes and facts link to the original sources.
Meta released Muse Spark Open, an open-weights model with 816 billion total parameters, 42 billion active per token, and a 1 million token context window. It natively supports multimodal input (text, image, video, audio) and is aimed at programming and agent tasks, especially long-running tasks, complex software engineering, and multi-step reasoning. Meta claims it is the most capable US open-weights model to date. The model runs on two Mac Studio devices with Apple M5 Ultra chips using the MLX framework. On the Artificial Analysis open-weights intelligence index, Muse Spark Open (xhigh) scored 39, below several Chinese open-weights models such as MiMo-V2.6-Pro (46), GLM-5.3 (45), Step 5 Preview, Kimi K3, and Ling-3.1-flash (41), and tied with DeepSeek V4.1 Flash. Developers can access the model via APIs from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, or download weights from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. It supports inference frameworks including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, and is optimized for AMD, Apple, Intel, and NVIDIA chips.
Muse Spark Open uses a sparse mixture-of-experts architecture with 816B total parameters and 42B active per token, enabling a 1M-token context window. It supports multimodal inputs and is optimized for multiple hardware vendors and inference frameworks, including Apple MLX for local execution on Mac Studio.
Meta's release puts a US open-weights model directly against Chinese open-weights models on public benchmarks, where it scores below several Chinese models on Artificial Analysis' intelligence index. Developers outside China now have a new open-weights option for coding and agent tasks, but must weigh its benchmark performance against Chinese alternatives.
The model is available for commercial use via API platforms and downloadable weights, but license terms are not specified in the evidence. Enterprises can evaluate it for coding and agent workflows, especially where local deployment on Apple hardware is desired.
A verifiable next signal is whether Meta releases detailed benchmark results for coding and agent tasks, and whether independent evaluations confirm the Artificial Analysis score. Another signal is adoption by developers on Hugging Face and API platforms, measurable through download counts and API usage.