Model releases & benchmarks
Kimi K3: Open Frontier Intelligence
Kimi K3 pairs 2.8T parameters with sparse expert activation and a 1-million-token context window; the launch case rests on agentic coding and knowledge-work demonstrations, with important evaluation caveats.
Why it matters: For long-running coding or research tasks, the examples suggest K3 may handle substantial work with limited oversight. Deployment is not plug-and-play: the provider recommends 64 or more accelerators, and warns that incompatible thinking-history handling can destabilize generation while ambiguous tasks can prompt excessive initiative.