Meta releases Muse Spark Open, an 816B open-weight model

An open-weight model that beats a vendor's own closed model on its own benchmarks does not come out every week.

Meta has released Muse Spark Open, an open-weight mixture-of-experts model with 816 billion total parameters and 42 billion active per token, published on Hugging Face with inference support in Transformers, vLLM, SGLang, llama.cpp and Apple's MLX. The context window is 1 million tokens, the model is multimodal on input (text, images, audio and video) and text-only on output, and Meta's blog post calls it "the most capable US open-weight model to date".

What Meta measures

The Hugging Face post describes the model as 62-layer with a hidden dimension of 8,192, 128 query heads and 16 key-value heads, 256 experts per MoE layer with 8 selected per token, and a 1.8 billion parameter perception encoder. Audio comes in at 12.5 tokens per second. Weights are published in BF16 and NVFP4, with MLX quants at 3 and 4 bits for Apple Silicon and GGUF quants for llama.cpp.

Meta's own figures put Spark Open ahead of Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra across the agentic and tool-use benchmarks the company reports, often by 20 points. Against Chinese open models, Spark Open leads Kimi K3, GLM 5.3 and Qwen 3.8 on JobBench and MCP-Atlas, and ties with them on SciCode. On coding, Kimi K3 scores 85.0 on Terminal-Bench 2.1 and GLM 5.3 scores 83.9, against Spark Open's 76.4. On DeepSWE 1.1 the spread is 68 to 60. These are Meta's own numbers against models it did not train, so a buyer should run its own agent on the models it is choosing between before trusting the chart.

Open weights, limited commercial terms

Meta releases Spark Open under what the company calls the Muse Model License. Internal use and building the model into a product are free, but any company with more than $20 million in gross revenue in any consecutive 12-month period "must notify Meta and enter into a separate written agreement with Meta before offering the Software or any Derivatives as part of that Model as a Service business". The licence also requires products built on the model to show "Built with Muse Spark Open" in the user interface, and bars military and critical-infrastructure deployments. The New Stack notes the licence is close in shape to the one Moonshot ships with Kimi K3.

The Apple partnership is a specific hardware pitch

The model is too large to run on a single consumer card, so Meta shipped it with an Apple partnership. Mark Zuckerberg wrote on X that Spark Open "runs great on new Mac Studios thanks to work with @johnternus and Apple", and the Hugging Face card gives a measured number: on an M3 Ultra Mac Studio with 512 GB of memory, the 2-bit quant runs at 18 tokens per second. Running the full 816B weights needs two of the new Mac Studios with M5 Ultra chips and 512 GB each, linked, with MLX managing the split. Apple does not currently sell that configuration to the public, and The New Stack notes that a pair of them will cost as much as a compact car when it does.

The model is reachable through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and the weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft's Foundry Model Catalog. For a team evaluating a frontier model to run inside its own account, Spark Open is a specific candidate with a licence that will apply at a specific revenue threshold.

Source

Welcome Muse Spark Open, Hugging Face blog, and Meta ships Muse Spark Open, its answer to China's best open-weight models, The New Stack.