Meta launches Muse Spark Open, an 816B open-weight AI model
Meta has released Muse Spark Open, its largest open-weight AI model to date, carrying roughly 816 billion total parameters and a context window of up to 1,048,576 tokens. The launch puts Meta Muse Open among the biggest openly available AI models in the world.
What Meta Muse Open brings to developers
The new model accepts text, images, audio and video as input and produces text as output. Meta built it for coding, tool use and agentic tasks.
Under the hood, Muse Spark Open uses a Mixture of Experts design. Each layer holds 256 experts, and about eight activate for every token processed, with around 42 billion active parameters per token. Only a fraction of the full network runs at any moment, which keeps the model efficient.
Meta is shipping the model in BF16 and NVFP4 checkpoint formats, alongside quantised versions for different hardware. On the audio side, it uses separate visual and audio encoders, with audio processed at 24kHz at roughly 12.5 tokens per second. The model supports tool calling and several reasoning levels, from low to high.
How the demos showed off its agentic skills
Meta paired the release with demos aimed at showing what the model can do on its own. In one, the model generated and improved 3D objects in Three.js, using browser screenshots to evaluate and edit its own code. In another, it controlled a browser purely from screenshots, without receiving any of the page’s HTML or DOM data. A third showed a voice-driven shopping task, where a user described a product and the model combined audio, visual and agentic abilities to find a match.
For context, see our coverage of Meta’s Muse gadgets and the intelligent UI updates expected in ChatGPT.
What open-weight actually means
The word “open-weight” does real work here. When a model is open-weight, anyone can download its trained weights and run the model on their own machines or servers. That differs from closed models, which live behind a company API queried over the internet.
Open weights matter for developers for three reasons. First, there are no per-query fees: once the model runs on your hardware, each question costs nothing extra. Second, data stays private because nothing leaves the local machine, which matters for firms handling sensitive records. Third, developers can fine-tune the model on their own data without asking anyone’s permission.
For developers in Pakistan, this opens a practical path: small teams can build AI tools without paying per-query charges to foreign API providers, and our look at artificial intelligence in Pakistan’s corporate sector shows demand is already growing. The IoT and AI roundup is another area where local developers could put open models to work.
Running it locally takes serious hardware
One catch is size. Even heavily compressed, Muse Spark Open needs more than 256GB of memory. It runs with llama.cpp, oMLX and mlx-lm on high-end Apple Silicon Macs: a 2-bit version tested on an M3 Ultra with 512GB unified memory used about 304GB and produced roughly 18 tokens per second, while a 3-bit version used about 344GB and reached about 21 tokens per second.
On the data-centre side, the BF16 version runs across multiple Nvidia H200 systems, and the NVFP4 format supports newer Blackwell GPUs such as the B200 and GB300. It also works with the SGLang and vLLM serving frameworks, plus Hugging Face Inference Endpoints.
Frequently asked questions
Source: ProPakistani
What is Meta Muse Open?
It is Meta’s largest open-weight AI model, with about 816 billion total parameters, roughly 42 billion active per token, and a context window of up to 1,048,576 tokens. It handles text, images, audio and video input and targets coding and agentic work.
What does open-weight mean?
Open-weight means the model’s trained files are publicly available, so anyone can download, run and fine-tune them. Closed models are only accessible through the owning company’s API.
How many experts does the model use?
Each MoE layer uses 256 experts, with eight activated per token. That keeps computation manageable despite the model’s large total size.
Can I run Muse Spark Open on a home computer?
Not a regular one. Even the most compressed versions need more than 256GB of memory, so local use requires a workstation like an M3 Ultra Mac with 512GB of unified memory or a multi-GPU server setup.
