AI Models

Meta Releases Open Muse AI Model for Coding and Agentic Tasks

Muse Spark Open has 816 billion total parameters, with 42 billion active parameters, and supports a context window of one million tokens.

Meta has released Muse Spark Open, an open-weight AI model designed for software engineering, complex reasoning and autonomous, multi-step tasks.

Muse Spark Open has 816 billion total parameters, with 42 billion active parameters, and supports a context window of one million tokens. The natively multimodal model is designed to handle long-running tasks, complex coding workflows and agentic applications.

In April, Meta unveiled Muse Spark, the first model in its new Muse family developed by Meta Superintelligence Labs.

"Muse Spark Open is available today through API platforms at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Model weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models," Meta said in a blog post.

Meta said the model can run on a desktop setup comprising two M5 Ultra Mac Studios using Apple's MLX machine-learning framework. Each system can provide up to 512GB of unified memory, allowing the two machines to accommodate the model's parameters. Meta describes the configuration as a way to bring advanced AI capabilities to local hardware.

And @AIatMeta went live with Muse Spark Open!

The 816B model was trained on NVIDIA GB300 NVL72. Its 1M-token context window gives coding agents room to work across large codebases.

Meta also produced NVFP4 checkpoints optimized for Blackwell Ultra, delivering over 50K… https://t.co/HJUKdC1n3D
— NVIDIA AI (@NVIDIAAI) October 15, 2026

The company demonstrated the model performing software development tasks through Muse Code, including fixing bugs, implementing features and building an iOS application.

It also demonstrated an agentic workflow in DaVinci Resolve, where the model reviewed video footage frame by frame to identify suitable clips and create vertical crops.

The model supports inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth. Developers can also customize it using PyTorch.

Meta said custom GPU kernels for SGLang deliver kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD GPUs, alongside a 3% end-to-end speed improvement on AMD MI350.

The company also highlighted safety measures incorporated into the model's post-training process. According to Meta, Muse Spark Open was evaluated under its Superintelligence Scaling Framework and met the required release thresholds across assessed risk categories.

The model is being released under Meta's Muse Model License, which includes limited commercial terms. The company said developers can further fine-tune it using PyTorch's TorchTitan framework.