What it does
Muse Spark Open is a model from Meta whose weights were released on October 15, 2026. It is natively multimodal, with 816B total parameters (42B active) and a context of 1M tokens. Meta positions it for coding and agentic work, such as extended multi-step tasks, software engineering and reasoning. There are several ways to use it. Hosted API access is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Supported inference engines include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth, with tuning for AMD, Apple, Intel and NVIDIA chips. Meta says it can be tuned further with PyTorch's TorchTitan. Meta highlights running it on a desktop using a pair of Apple M5 Ultra Mac Studios with MLX, each machine offering up to 512 GB of memory to its GPU. The weights come under the Muse Model License, whose limited commercial terms the page says make the model free for many developers. The announcement also says the model received additional safety post-training for its open release.
Capabilities
- Natively multimodal, with 816B total parameters and 42B active
- Context window of 1M tokens
- Aimed at coding, extended agentic tasks and multi-step reasoning
- Hosted APIs from partners including Together AI, Fireworks AI, Groq and OpenRouter
- Weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models
- Can run locally with MLX on a pair of Apple M5 Ultra Mac Studios
When to use it
- Coding and software engineering agents that need a very long context
- Running a large model on your own hardware or infrastructure
- Customizing an open-weight model with PyTorch
- Trying the model through a hosted API without downloading weights
