Meta released Muse Spark Open, an open-weight coding/agent model
Meta AI Research ● Covered by 25 sources
Meta released Muse Spark Open, an open-weight coding model with a 1M-token context window. It can run on desktop Macs and is meant for serious agent work, not demos.
Based on reporting by Meta AI Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta has released Muse Spark Open, its latest open-weight model for coding and agent-style work. The company is positioning it as the most capable U.S. open-weight model it has shipped so far, and it’s available now through several hosted platforms and for direct download.
The headline here isn’t just size, although the size is hard to ignore. Muse Spark Open is a natively multimodal model with 816B total parameters, 42B active, and a 1M-token context window. Meta says it is built for long-horizon tasks, complex software engineering, and multi-step reasoning. That’s the pitch: not a chat toy, but something that can stay on task across a sprawling job.
The other hook is where it can run. Meta says Muse Spark Open is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework. The company says Apple’s unified memory and bandwidth make the setup work, and that two machines can hold the full model. Meta also says it tested the model doing on-device agent workflows, including shipping real iOS changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve end to end.
This is also very much a distribution story. The model is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The weights are on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. It runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon.
Meta says the open-weight release comes with extra safety post-training and that the model met the thresholds required under its Meta Superintelligence Scaling Framework. It’s releasing the weights under the Muse Model License with limited commercial terms. So this is open, but not open in the romantic, anything-goes sense. Meta is clearly trying to thread the needle: broad access, controlled terms, and a model it wants developers to actually use.
My take — AI-written commentary, not fact-checked reporting
Open weights keep winning the real argument here: useful models should be runnable, tunable, and not trapped behind a rent bill. Meta is also doing the classic big-platform move of calling that freedom while keeping the fine print close at hand. Still, if the choice is between another glossy demo and something that can edit iOS code on a desktop, the desktop wins.
Read more about this at: Meta AI Research
