SourcesKili Technology
publisher · 1 source
Kili Technology
1 source read from this publisher, newest first, with the methods each one describes.
DeepSeek V4 Guide: Engram Memory, Training Data Strategy & Release Status (2026)
news analysis · third party analysis · 2026-03-23 · 33 methods
Mixture of expertsThinking modeSupervised fine-tuningReinforcement learningDeepSeek Sparse AttentionGroup Relative Policy OptimizationMuonManifold-Constrained Hyper-ConnectionsOn-Policy DistillationGenerative Reward ModelCompressed Sparse AttentionMid-trainingHeavily Compressed AttentionMXFP4 Quantization-Aware TrainingBatch-invariant deterministic kernelsEngramFull-vocabulary logit distillationAnticipatory RoutingFiltering batched auto-generated and templated web contentInternal engineering-task evaluationQuick Instruction tokensSample-level attention maskingSwiGLU clampingXML-based tool-call schema with DSML tokenLong-document curationMultilingual corpus expansionPairwise comparison for Chinese writingPreemptible rollout service with write-ahead loggingRubric-guided RL dataSandbox infrastructure (DSec)Specialist trainingTask-specific data injectionWhite-collar enterprise-task evaluation