Meta opens its big AI model, but Chinese rivals score higher in coding

On October 15, 2026, Meta released the weights for its Muse Spark Open language model. The weights are what the model has learned, and once they are open, developers can download the model from Hugging Face and run or customize it themselves. It is released under Meta's own license, the Muse Model License.

Muse Spark launched as a closed model on April 8, 2026, and Meta has since promised an open version. It is Meta's first frontier model without open weights. According to Meta, the new version can be used for coding and agent work, meaning tasks where the AI solves multiple steps in sequence on its own. It accepts text, images, video and audio as input, but responds only with text.

Large model with lots of memory

The model has 816 billion parameters in total, but uses only 42 billion active parameters per token. That is because it is a mixture-of-experts model, where only part of the model works at a time. The context window, meaning how much text it can keep track of at once, is 1 million tokens.

Meta calls it the most capable American open-weights model to date. That is Meta's own assessment. The model is not Meta's most advanced: Muse Wonder remains closed.

Behind Chinese models in coding

According to Meta's own comparison table, Muse Spark Open scores 76.4 points on the Terminal-Bench 2.1 coding test. Kimi K3 scores 85.0, GLM 5.3 scores 83.9 and Qwen3.8 scores 82.0. On MCP-Atlas, which measures tool use, the Meta model scores 89.2, the highest in the table, versus 84.5 for Qwen3.8. On the Artificial Analysis Intelligence Index, it scores 39 versus 44 for Kimi K3, 45 for GLM 5.3 and 40 for Qwen3.8.

The figures come from Meta's own table and had not been independently verified at launch. The full evaluation report had not been published. Short tests also do not necessarily show how the model handles long tasks.

Terms, pricing and hardware

The license is free for most developers. Companies with more than USD 20 million in gross revenue over 12 months, including affiliates, must notify Meta and enter into a separate hosting agreement. Use within a company's own operations and embedding in its own products are exempt. Users must display "Built with Muse Spark Open," and military use and critical infrastructure are prohibited.

The model is large: Meta's BF16 version is about 1.6 TB, and the NVFP4 version about 530 GB. Smaller versions also exist from AMD, Intel and Unsloth, among others. It is hosted by Together AI, Groq and OpenRouter, among others, and is available in catalogs at AWS, Google Cloud and Microsoft. At Together AI, it costs USD 1.25 per million input tokens and USD 4.25 per million output tokens.

Meta and Apple have tested the model on Mac Studio. Two M5 Ultra machines with 512 GB deliver 24.99 tokens per second, and four with 256 GB deliver 36.73. The price and release date for those machines are not stated in Meta's own sources.