Meta's Muse Spark Open: 816B, free weights

Meta released Muse Spark Open, a model with 816 billion parameters that anyone can download, on the 15th of October 2026. Meta called it "the most capable US open-weight model to date" in its research blog.

Parameters are the values a model learns in training. Open weights mean the trained model file is public, so companies can run it on their own computers.

816 billion parameters, 42 billion at work for each word

Muse Spark Open is a mixture of experts, Hugging Face's launch post of the 15th says. Each layer holds 256 small expert networks, and 8 of them handle each token.

So about 42 billion parameters do the work for each token, Meta's blog says. That is about 5 percent of the model, by our arithmetic.

The model reads text, pictures, sound and video, and writes text, Meta's developer docs say. It can hold 1 million tokens, small pieces of text, in one go.

Below we break down the scores in Meta's own table

Meta's model card compares Muse Spark Open with seven open models. It leads MCP-Atlas, a test of tool use, at 89.2, ahead of Alibaba's Qwen3.8 at 84.5.

On Terminal-Bench 2.1, a coding test, it scores 76.4. Moonshot AI's Kimi K3 scores 85.0 and Z.ai's GLM 5.3 scores 83.9 in the same table.

Two Mac Studios with 512 GB each run it

Meta said the model runs on two Apple Mac Studios with M5 Ultra chips. Each has 512 GB of memory, so the pair works like one machine with about 1 terabyte.

The full weights take about 1.65 terabytes, Meta's docs say. The 4 bit copy takes about 537 GB and fits one server with four NVIDIA Blackwell chips.

Free for most, a deal with Meta above 20 million dollars

The Muse Model License lets companies use the model for free, with two conditions. Products must show "Built with Muse Spark Open" somewhere users can see it.

Big sellers need more. Firms selling it as a service with over 20 million dollars in yearly sales must sign a deal with Meta first, the license says.

Key facts

Released by Meta on the 15th of October 2026, weights on Hugging Face.

816 billion parameters, 42 billion active, 1 million tokens of context.

Leads MCP-Atlas at 89.2, trails Kimi K3 on Terminal-Bench 2.1, 76.4 to 85.0.