On the 15th local time, Meta launched the open-weights model Muse Spark Open, loudly proclaiming it the most capable American open-weights model to date. It adopts a natively multimodal architecture with 816B total parameters and 42B active parameters, a context window stretched to 1M tokens, and is honed for code building and agent collaboration, showing outstanding performance on long-horizon tasks, complex software engineering, and multi-step reasoning.
The most eye-catching part is that the barrier to local deployment has been pushed extremely low: relying on the latest MLX framework, it becomes the first model in its class that can run locally on the desktop with only two Mac Studios equipped with M5 Ultra chips (512GB memory version). In other words, an affordable workstation setup is enough to support a top-tier open-source large model, without having to depend on cloud data centers.
Rolled out across all major platforms, weights free for commercial use
Starting today, developers can access it through API platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI; the model weights have also landed for public download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. On the inference side, it is compatible with mainstream engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, has been deeply optimized for AMD, Apple, Intel, and NVIDIA chips, and can be conveniently further developed based on PyTorch.
In its announcement, Meta emphasized that the model's safety alignment was strengthened during post-training, with benchmarks showing tight safeguards across all risk dimensions. The model is released under the Muse Model License, which comes with restrictive commercial terms, and most developers can obtain and use it for free. At a time when OpenAI and Anthropic are tightly guarding their weights, Meta hanging the "America's strongest" sign on the door of open source is both a display of technical muscle and an effort to seize discourse power over the developer ecosystem.