It looks like Meta has put up an 816 billion parameter model on Hugging Face and encouraged everyone to take what they could use. More or less. Muse Spark Open was launched on October 15, and Meta says it is the most powerful US open-weight model ever. That’s a bold claim but not necessarily the last word on the matter. Benchmarking is done by Meta.
Also read: Maya 2 explained: A free open weight voice AI model for 22 Indian languages
Muse Spark Open is inherently multimodal, having 816B total parameters where 42B parameters are live at any time, and a 1M-token context window. The live parameters count does make sense. Not all parts of the model are fired for every single token, and that’s what makes it work. Meta says that the model is great for long-horizon tasks like software engineering and multi-step reasoning.
To demonstrate its skills, Meta performed tests with its Muse Code application with respect to the whole model deployed on-device. It delivered actual iOS updates to Ice Cubes, a completely free Mastodon client, implementing some bugfixes and new features in a complete build. It powered DaVinci Resolve and did the review of video footage frame by frame before cropping its best parts into vertical videos. Imagine that, video footage review by the model on your desk.
Meta claims that Muse Spark Open is the only model in its category to function on a desktop platform, through two Mac Studios powered by Apple’s M5 Ultra chip, and the new MLX framework. Both machines have access to 512GB of unified memory and 1.2TB/s of memory bandwidth, meaning that both machines are capable of working like one single computer with a terabyte of memory.
Meta tries to portray this as the ability to work with advanced models from anywhere, including your house. However, now for the harsh truth. Two premium Mac Studios cost money similar to that of a workstation, and this amount in rupees is substantial. Hence, “runs at home” is really a well-funded house.
Hosted APIs are up and running on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Pretrained weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. The model executes in SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with hardware optimizations for AMD, Apple, Intel and NVIDIA chips.
Meta delivered their own Helion GPU kernels in SGLang, with improvements of 1.03x to 3.95x of kernel performance on NVIDIA and AMD GPUs. The total speedup on AMD MI350 GPU is only 3%, which is rather small. Kernel speeds sound more impressive on a slide than end-to-end speedup does.
The weights are made available under the Muse Model License with restricted commercial provisions. Meta claims that many developers can leverage the model without paying any fees. Many, but not all developers. The weights are open but not the source code. And that makes a difference once you build your project using it. Read the license.
As the weights have been made open, Meta claims to have incorporated additional safety mechanisms after the training process and the model was also tested according to its own Meta Superintelligence Scaling Framework. The model card is already live on Hugging Face and the detailed evaluation report will be released shortly. Till then, the claim remains on Meta’s head.
Also read: Amazon Bee AI wearable is here: What it does and why everyone is talking about it