Image: Meta
Meta has released Muse Spark Open, a free-to-download version of its Muse Spark AI model that anyone can run on their own hardware. Meta calls it "the most capable US open-weight model to date" and says it is built for coding and agent-style work that takes many steps.
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
— AI at Meta (@AIatMeta) October 15, 2026
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token… pic.twitter.com/yAtoxNeYJm
"Open-weight" means Meta is handing out the trained model itself, not just access through an app. Developers can download it, fine-tune it and build products on it without sending data to Meta.
The model is big. It has 816 billion parameters in total, but only 42 billion are used for each word it produces, which keeps it faster and cheaper to run than its size suggests. It can take in up to 1 million tokens at once, well over a thousand pages of text, and it understands text, images, video and audio, though it only replies in text.
Meta's own numbers are strong in places. In the benchmark table it published, Muse Spark Open leads on agent tests such as JobBench (61.8) and MCP-Atlas (89.2), and on the tough HLE-Diamond reasoning test without tools (26.1).
Image: Meta
But read the same table closely and the "US" in Meta's claim matters. On coding, the job Meta says the model was built for, China's Kimi K3 and GLM 5.3 beat it on both Terminal-Bench 2.1 and DeepSWE 1.1. On the overall Artificial Analysis Intelligence Index listed in the model card, Muse Spark Open scores 39, behind GLM 5.3 (45), Kimi K3 (44) and Qwen3.8 (40).
So the fair reading is that Meta now has the strongest American open model, ahead of rivals such as Nvidia's Nemotron 3 Ultra and Reflection AI's Beam, while Chinese labs still hold the overall open-model lead. All of these scores come from Meta and haven't been independently checked yet.
Running it yourself is not a laptop job. The model card says it needs at least 700GB of graphics memory in its compressed NVFP4 version and 2TB at full precision. Meta's home option is two Apple Mac Studios with M5 Ultra chips, each with up to 512GB of memory, linked using Apple's MLX software.
For most people, the easier route is a hosted service. Meta says it is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hugging Face downloads require an access request that Meta approves.
The licence is generous but not fully open. The Muse Model License lets people use, modify and sell products built on the model for free, but they must show "Built with Muse Spark Open." Companies earning more than $20 million a year that resell AI models through an API need a separate deal with Meta. The model is also not meant for anyone under 18.
Meta says the open version got extra safety training because it ships without the guardrails of Meta's own apps, and that it passed the risk checks in its Superintelligence Scaling Framework. A full safety evaluation report is still to come.
Comments