Meta is back in the open-weight race
Meta has released Muse Spark Open, a natively multimodal model with 816 billion total parameters and 42 billion active per token, a context window of one million tokens, and open weights on Hugging Face. Meta calls it “the most capable US open-weight model to date” and aims it at coding and agentic work — long-horizon tasks, multi-step software engineering, and the kind of job that runs for hours rather than seconds.
It is the company’s first serious open release since the Llama line, and it arrives into a field where the strongest downloadable models have been Chinese. Meta positions it against Qwen 3.8, GLM 5.3 and Kimi K3, and The Decoder reports that Meta’s own figures put it ahead of those three in agentic, knowledge-and-reasoning and multimodal categories, behind some of them on coding, and well short of Meta’s paid Muse Wonder Full models. Those are vendor numbers; the full evaluation report is, in Meta’s words, available soon.
The sparsity is the product
An 816-billion-parameter model that only lights up 42 billion per token is a model built to be run by people who do not own a data centre. Meta says the pair of Apple M5 Ultra Mac Studios it demonstrated — each with up to 512GB of unified memory and 1.2TB/s of bandwidth, acting as one machine with roughly a terabyte — ran the model locally through the MLX framework. The demonstrations were agentic rather than conversational: fixing bugs and adding features to Ice Cubes, an open-source iOS Mastodon client, and reviewing and cropping footage in DaVinci Resolve.
The serving work is spread wide. Meta says custom Helion GPU kernels in SGLang deliver kernel-level speedups between 1.03 and 3.95 times on Nvidia and AMD hardware, and a 3% end-to-end gain on AMD’s MI350. Inference runs through SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and Nvidia. Weights are also on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted APIs come from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Fine-tuning goes through PyTorch’s TorchTitan.
“Open” with conditions
The weights come under the Muse Model License, which Meta describes as carrying limited commercial terms that let many developers use the model free of charge. The Decoder reports two specifics: anything built on it has to carry the label “Built with Muse Spark Open”, and providers with annual revenue above $20 million who want to resell it as an API service need a separate agreement with Meta. That is the familiar shape of a Meta open release — downloadable, modifiable, and not quite open source.
On safety, Meta says the model received additional post-training for open-weight risks and met the release thresholds in its Meta Superintelligence Scaling Framework.
What to watch
The evaluation report Meta has promised is the thing to wait for, because every comparison published so far is Meta’s own. The second test is whether anyone outside Meta reproduces the two-Mac-Studio demonstration at a speed that makes it more than a proof of concept.
