SourcesKimi Team
publisher · 1 source
Kimi Team
1 source read from this publisher, newest first, with the methods each one describes.
k3 tech report
technical report · 154 methods
Multi-Teacher On-Policy DistillationSupervised fine-tuningMuonQuantization-Aware TrainingLatentMoEPreserved thinking historyReasoning token-budget controlAttention ResidualsStable LatentMoEKimi Delta AttentionCosine decayNative multimodal architecturePartial rolloutQuantile BalancingWarmup-Stable-DecayAuxiliary-loss-free load balancingHyperparameter scaling-law fittingMoonEPMXFP4 Quantization-Aware TrainingNo Position EncodingPer-Head MuonFunction callingGated Multi-head Latent AttentionSiTU-GLUBlock Attention ResidualsCo-located rollout and trainingKDA-aware prefix-cache managementProgressive context extensionSynchronization-free static-shape MoE executionZero-copy fused token permutation and unpermutationAgent-driven recursive knowledge graph constructionAgentic Generative Reward ModelAgentic judge protocolAgentic visual reasoning with a Python interpreter in a sandboxautonomous chip design and verification with open-source EDA toolsAutonomous Execution TasksBlind expert judgingBlock-wise FP8 activation quantization with offloadBudget-based verbosity controlCache-based pipeline communication