PickMeta releases Muse Spark Open weights
Meta released Muse Spark Open, an 816B-parameter open-weight multimodal model (42B active) with a 1M-token context window for coding and agents. Weights are on Hugging Face; hosted APIs include OpenRouter, Groq and others, and it can run on two Apple M5 Ultra Mac Studios via MLX.
Meta has released the weights for Muse Spark Open, which it describes as the most capable open-weight model from a US company so far. It is a large, natively multimodal model aimed squarely at coding and agentic work, and it can be downloaded, self-hosted and fine-tuned.
The specs
Muse Spark Open uses a mixture-of-experts design with 816 billion total parameters, of which about 42 billion are active for each token. It supports a context window of one million tokens. Meta says the model is tuned for long-horizon tasks, complex software engineering and multi-step reasoning, the kind of work agents do when they run for a long time on a large codebase.
Where you can run it
Distribution is broad from day one. Hosted APIs are available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights themselves can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta lists support for common inference engines including vLLM, SGLang, TensorRT-LLM, Transformers, llama.cpp, MLX and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA hardware.
The most striking claim is about local use. Meta says Muse Spark Open runs on two Apple M5 Ultra Mac Studios using the latest MLX framework. Each machine can give its GPU access to a large pool of unified memory, so together they can hold the whole model, and because only a small slice of parameters is active per token, memory bandwidth is enough to generate at usable speed. Meta showed on-device agent demos, including making real changes to an open-source iOS app.
Why it matters for builders
Open weights change the economics of building with frontier-class models. You can pick the cheapest provider, switch providers without rewriting prompts, keep sensitive code on your own hardware, and fine-tune for your domain. A one-million-token context also makes it realistic to feed entire repositories or long document sets to an agent without heavy retrieval plumbing.
The trade-offs are familiar: running an 816B model yourself still requires serious hardware, and open weights shift responsibility for safety and misuse controls onto whoever deploys them. For most teams, the practical path is to start on a hosted API such as OpenRouter or Groq, measure quality on your own tasks, and only then consider self-hosting.
You can find Muse Spark Open and the platforms that serve it in our New AI Models and APIs categories.
Source: Meta AI Research. This article is PickAI’s own summary and analysis based on the original reporting.
