Meta has released a giant open-weight model. On October 15, 2026, Meta announced Muse Spark Open, an AI system with 816 billion total parameters, multimodal input support, and a 1M-token context window.
Meta says Muse Spark Open is the most capable US open-weight model to date. Meta’s chief of AI Alexandr Wang wrote on X: “is the most capable US open-weight model to date.” The claim is clear, and so is the scale behind it—this is not a compact model wearing an ambitious press release.
Meta Superintelligence Labs released Muse Spark Open with an architecture that includes 62 layers, a hidden dimension of 8192, and a vocabulary containing 202,048 tokens. The model also supports 42 billion active parameters per token, alongside its 816 billion total parameters.
A Huge Model With Consumer Hardware Claims
The release stands out because Meta says Muse Spark Open can run on consumer hardware. The company used two Apple M5 Ultra Mac Studios for deployment, with each Mac Studio carrying 512GB of memory.
That setup provides 1TB of combined memory across the two Mac Studios, though the listed hardware requirements remain demanding. Running Muse Spark Open with NVFP4 requires at least 700GB of VRAM, while BF16 requires 2TB of VRAM.
Those figures place the model in an unusual category: open-weight access paired with hardware requirements that reach far beyond a typical personal computer. The consumer-hardware claim is real, but “consumer” covers a broad neighborhood when two high-memory Mac Studios are part of the address.
Context Length And Memory Requirements
The 1M-token context window is another defining feature of Muse Spark Open. That context length gives the model a large token capacity, while the architecture’s 62 layers and hidden dimension of 8192 define the technical structure Meta disclosed for the system.
Memory remains the practical limit. The listed requirement of at least 700GB of VRAM for NVFP4 and 2TB for BF16 shows how the model’s total parameter count affects deployment, even with only 42 billion active parameters per token.
Unsloth also provides 300GB setups for Muse Spark Open. That figure offers another route for running the model, but it does not erase the scale of the system’s memory demands. Open weights remove one gate; they do not make hardware free.
Muse Spark Open supports multimodal inputs, giving it capabilities beyond a model limited to text. Meta has tied that support to an open-weight release, a 1M-token context window, and deployment across high-memory hardware configurations.
The announcement therefore combines two messages. Meta is presenting Muse Spark Open as the most capable US open-weight model to date, while its deployment figures show the cost of reaching that level: 816 billion total parameters, large memory requirements, and specialized configurations even when the hardware remains commercially available.
For developers and researchers, the important details are concrete rather than mysterious. Muse Spark Open offers open weights, multimodal inputs, a 1M-token context window, and multiple deployment paths, including two Apple M5 Ultra Mac Studios and 300GB setups via Unsloth.
The model’s release does not make large-scale AI hardware disappear. It makes the boundary more visible—open access on one side, and hundreds of gigabytes or terabytes of memory on the other.
Based on
- Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context — marktechpost.com
- Meta unveils Muse Spark Open, the new top U.S. open weights model — and it runs on two paired Mac desktops | VentureBeat — venturebeat.com
- Meta releases Muse Spark Open, a new ‘open-weight’ AI model | AI (artificial intelligence) | The Guardian — theguardian.com
