A trillion total parameters sounds like a deployment barrier, but the more useful number here is roughly 49B active per token: the sparse MoE design aims to combine broad model capacity with a smaller active computation footprint. The trade-off is that hosting the full checkpoint still demands substantial hardware; sparse activation does not make the weights small.
Key Capabilities
- Long-context work: A 1M-token context can keep large collections of documents or extended task histories in scope, though practical throughput and memory depend on the serving setup.
- Text-and-image input: It handles images, charts, and documents alongside text, while producing text output; this is useful when visual evidence must inform a written answer or tool-driven workflow.
- Reasoning and tool use: Configurable reasoning effort, function calling, and strict JSON output let the same model support both direct responses and structured agent workflows.
- Sparse expert routing: Four of 128 routed experts are selected per token, alongside a shared expert. This supports specialization without activating every expert for every token.
Who It Suits
Great fit if you need one open-weight model for long-context, multilingual, visual, or tool-using workloads and have infrastructure suited to a very large checkpoint. Look elsewhere if you need lightweight local inference or image generation: this model accepts images but returns text, and its open-weight license requires a separate commercial license for Model-as-a-Service businesses.
