The model uses 512 routed experts with ten routed experts plus one shared expert active for each token. It has a native 262K-token context window extensible to about 1.01 million tokens, and it uses gated linear attention and gated attention blocks to manage large-scale inference.
The Hugging Face release includes instructions for vLLM, SGLang, Docker Model Runner, and quantized local runtimes. It is useful for organizations with substantial inference infrastructure that need high-capacity open model weights for long-context agent workloads.
