A useful distinction in large models is the gap between total capacity and the compute used for each token. Here, 594 billion total parameters coexist with 17 billion active parameters, letting the model draw on a broad expert pool without activating every expert for every step. That trade-off helps explain its focus on long-running agent work, while also making deployment hardware a major consideration.
Key Capabilities
- Text and image input, text output: The shared multimodal decoder can handle visual questions and document-oriented tasks alongside ordinary language work; it does not generate images or process audio.
- Sparse expert routing: Each token is routed to 8 of 384 experts across a 48-layer decoder. This provides substantial model capacity with fewer active parameters per token than a dense model of comparable total size.
- Long context and agent workflows: A context window of up to 1 million tokens is suited to tasks involving extensive codebases, documents, or extended tool interactions. The model is also intended for coding assistants, retrieval-augmented systems, and tool-using agents.
- Open-weight deployment options: Released under Apache 2.0, it can be used for research, fine-tuning, and downstream applications. The native NVFP4 checkpoint is the recommended option; the alternative MXFP8 checkpoint requires more aggregate GPU memory for only a marginal reported evaluation gain.
Who It Suits—and the Trade-off
A good fit for teams evaluating an open-weight multimodal model for coding, visual understanding, or long-horizon agent tasks, especially when they have access to substantial GPU infrastructure. Look elsewhere if you need audio input, modest local hardware requirements, or a polished conversational style: the model can be verbose, may choose tools poorly in complex skill-heavy codebases, and can hallucinate or lose reliability in long conversations. High-stakes outputs still require domain validation and human review.
