Key Highlights:
- Meta rolled out Muse Spark Open, an 816-billion-parameter open-source AI model with 42 billion active parameters and a 1-million-token context window.
- The model is designed for coding, complex software engineering, and long-running agentic tasks.
- Muse Spark Open is available through several API platforms and model repositories, with support for leading inference engines and hardware from AMD, Apple, Intel, and NVIDIA.
Meta has released Muse Spark Open, an open-weight AI model designed to handle advanced coding and agentic tasks while giving developers more flexibility in how they run AI systems. The model has 816 billion total parameters, 42 billion of which are active, and a 1-million-token context window. Meta says it is the most capable US open-weight model to date and has made its weights available through multiple platforms. One of the launch’s key features is its ability to run on a desktop setup using two Apple M2 Ultra Mac Studios, bringing a model of this magnitude to locally operated hardware rather than relying on cloud infrastructure.
The release also expands developers’ options for customization, deployment, and inference, with support for several AI frameworks and hardware platforms. Meta says the model includes additional post-training safety and was assessed under its Superintelligence Scaling Framework before release.
What Makes Meta’s Muse Spark Open Different From Other Open-Weight AI Models?
Muse Spark Open is built for coding and agentic work, focusing on long-horizon tasks, complex software engineering, and multi-step reasoning. Its architecture combines 816 billion total parameters with 42 billion active parameters, while its 1-million-token context window is designed to accommodate extensive inputs for tasks that require processing large amounts of information.
Meta describes Muse Spark Open as the most capable US open-weight model available to date. Its published benchmark scorecard compares the model with several other systems, including Inkling, Nemotron 3 Ultra, Beam, Mistral 4 Large, Qwen3.8-2.4T-A95B, GLM 5.3 and Kimi K3. The assessments cover agentic capabilities, knowledge and reasoning, coding, and tool-assisted multimodal tasks. It doesn’t provide individual benchmark scores in its accompanying text, directing readers to the scorecard and assessment methodology for details.
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date. 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 context window → Natively multimodal across text, image, video, audio
— AI at Meta (@AIatMeta) October 15, 2026
A major factor of the rollout is its optimization for Apple Silicon. Using the latest version of Apple’s MLX framework, Meta optimized the model to run on two Mac Studios powered by M5 Ultra chips. Each machine can provide up to 512GB of unified memory, allowing the combined setup to accommodate the model. Meta also points to the chips’ 1.2TB/s memory bandwidth as a factor in handling inference memory demands.
The firm tested several agentic workflows running entirely on-device through Muse Code, which is powered by Muse Spark Open. These included implementing changes in Ice Cubes, an open-source Mastodon client, such as fixing bugs, adding features, and completing a build. Another demonstration involved controlling DaVinci Resolve, where the model reviewed footage frame by frame to identify strong moments and crop them into vertical formats.
Where Can Developers Access Muse Spark Open and What Safety Measures Does Meta Include?
Muse Spark Open is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Developers can also download its weights through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.
The model supports several inference engines, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Meta says the release includes custom, auto-tunable Helion GPU kernels in SGLang, delivering kernel-level speedups of 1.03 to 3.95 times on Nvidia and AMD GPUs, alongside a 3% end-to-end speedup on AMD MI350.
Developers can further customize the model using PyTorch’s TorchTitan. For Apple Silicon deployments, Meta recommends MLX, which can pull the model weights directly from Hugging Face. The company is releasing the weights under the Muse Model License, which includes limited commercial terms and allows many developers to use the model for free.
Safety is another part of the rollout. Meta says open-source models raise distinct considerations because users can download and run their weights outside the company’s infrastructure. In response, Muse Spark Open includes additional post-training safety. Meta says its assessments covered the most severe risks under the Meta Superintelligence Scaling Framework and that the model met the framework’s release thresholds.
The firm has linked to the Muse Spark Open model card for detailed results and said it will provide a full assessment report. With Muse Spark Open, Meta is combining open-weight access with support for cloud APIs, several hardware platforms and local desktop deployment. Its roll out gives developers another choice for building coding and agentic applications. Its licensing terms and guardrails will remain key considerations for conglomerates deciding how to use the model.
Also Read: Meta Says Muse Spark 1.3 Beats GPT-5.6 Sol at Coding
