Kimi Team: persistent long-horizon agent-training infrastructure

What changed

Integrated balanced expert execution, memory-aware hybrid-model kernels and persistent rollout/sandbox states to train long-horizon agents with million-token trajectories in Kimi K3.

Scientific scope

Integrated balanced expert execution, memory-aware hybrid-model kernels and persistent rollout/sandbox states to train long-horizon agents with million-token trajectories in Kimi K3. System-level integration beyond K2, not invention of KDA, Attention Residuals, latent experts or on-policy distillation. Model size alone is not the contribution. Reported scaling/agent outcomes are not independently reproduced; release license is model-specific.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. System-level integration beyond K2, not invention of KDA, Attention Residuals, latent experts or on-policy distillation. Model size alone is not the contribution. Reported scaling/agent outcomes are not independently reproduced; release license is model-specific.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Kimi K3: Open Frontier Intelligence
• Kimi Team: persistent long-horizon agent-training infrastructure — official supporting account

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction