Meta officially released the large-scale AI model Muse Spark Open weights on October 15, claiming that this is currently the most powerful open weight model in the United States, focusing on program development, multi-step reasoning and long-term execution of AI agent tasks. This model has 816 billion parameters, supports 1 million token contexts, and is optimized for the M5 Ultra chip in cooperation with Apple, allowing developers to run it locally through two Mac Studio without relying entirely on cloud GPU services.
Muse Spark Open uses a mixed expert (MoE) architecture with a total parameter count of 816 billion (816B), but approximately 42 billion parameters are enabled each time a token is processed to reduce the amount of calculation required for inference. The model natively supports text, image, video and audio input, and the output is mainly text. It can be used to analyze large code libraries, operate development tools, and process workflows that require repeated execution and modification. This model is based on Muse Spark 1.2, which has previously not opened weights, and adds additional security post-training, allowing developers to obtain a certain degree of security protection when deploying on their own. Meta stated that the model has passed the release threshold of the internal security framework, but it is still recommended that enterprises establish additional tool permissions, data access and content security mechanisms based on actual uses.
This release also brings out the cooperation between Meta and Apple on AI model deployment. According to official instructions, Muse Spark Open can take advantage of the latest version of the MLX framework and run on two M5 Ultra Mac Studio equipped with 512GB of unified memory each. The two computers collaboratively process model inference through high-speed connections, allowing the weights of large models to be stored in a distributed manner, eliminating the need for a single device to bear all the calculations.
This type of deployment still requires high-end hardware. The BF16 weight capacity provided by Meta is about 1.65TB, and the quantized NVFP4 version is about 530GB, which cannot be fully loaded by ordinary consumer-grade computers. Developers can also switch to GPU servers such as NVIDIA and AMD, or deploy through SGLang, vLLM, MLX and other inference frameworks. Meta also provides Helion optimization core for GPU inference, achieving acceleration effects in some computing links.
The evaluation results of Muse Spark Open show that the model has advantages in tool invocation of AI agents. In the MCP-Atlas test published by Meta, Muse Spark Open scored 89.2 points, higher than the Qwen, GLM and Kimi models compared in the same field; JobBench scored 61.8 points, slightly higher than the 61.4 points of GLM 5.3. These tests mainly observe whether the model can understand the work goals, select appropriate tools, and complete subsequent steps according to the execution results.
Programming capabilities show different results. In the Terminal-Bench 2.1 test, Muse Spark Open scored 76.4 points, lagging behind Kimi K3's 85.0 points, GLM 5.3's 83.9 points, and Qwen3.8's 82.0 points. Foreign media "SiliconANGLE" also pointed out that Meta's new model has outstanding tool usage capabilities, but when it comes to pure program development tasks, China's open weight model still has a competitive advantage. Relevant numbers come from comparative tests published by Meta. Actual performance will still be affected by inference settings and agent execution environment.
Model weights are currently available on Hugging Face, with BF16 and NVFP4 versions available. Developers can download and deploy or fine-tune themselves. Users who do not want to prepare their own hardware can also call it through platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry are also included in the model provisioning pipeline.
In terms of licensing, Muse Spark Open adopts Meta's customized Muse Model License, which allows developers to use, modify and create derivative models for free, but it is not without commercial restrictions. Products that use models must be marked "Built with Muse Spark Open" in accordance with regulations; if a company provides model hosting services, and the total revenue of the company and its affiliated companies exceeds US$20 million for 12 consecutive months, an additional licensing agreement must be signed with Meta.
Meta launched the smaller Muse Glimmer in August this year, with a parameter size of about 30 billion, mainly for native AI applications. The opening of a large version of Muse Spark this time further allows developers to obtain model weights suitable for complex agent tasks. The two can also be deployed by themselves, but the required hardware gap is quite large. For teams that lack high-end GPUs or large-capacity unified memory devices, Muse Spark Open can still be used through the managed API at this stage without having to set up a complete inference environment by themselves.
