SourcesNVIDIA Technical Blog
publisher · 2 sources
NVIDIA Technical Blog
2 sources read from this publisher, newest first, with the methods each one describes.
NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning | NVIDIA Technical Blog
first party release · official blog · 2026-10-15 · 30 methods
Mixture of expertsMulti-Token PredictionMulti-Teacher On-Policy DistillationSupervised fine-tuningReinforcement learningSpeculative decodingLow-Rank AdaptationPrefix cachingMTP BoostingPost-training quantizationQuantization-aware distillationTwo-stage supervised fine-tuningDisaggregated servingHybrid architectureInternal prefill checkpointingKV-cache-aware routingLiLiCorrMamba checkpointing for speculative-decoding replaySession affinity routingSkew reverse-KL distillationAttention-cache and Mamba-state preservationComponent-wise precision selectionContext reuseContinued pretraining with natural and synthetic dataDomain-specific post-trainingKV-cache managementKV-cache offloadingMixed-precision inferenceNVFP4 weight-and-activation quantizationSynthetic data generation with NeMo Data Designer
NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents | NVIDIA Technical Blog
first party release · official blog · 2026-08-11 · 15 methods
Mixture of expertsMulti-Token PredictionSpeculative decodingDSparkSparse expert activationDFlashMTP BoostingNVFP4 quantizationArtificial Analysis Intelligence IndexConcurrency-aware draft-length tuningfrontier/execution model division of laborHarness-optimized trainingModel routing and orchestrationModel routing by planning and execution roletask-to-best-model routing with NeMo Switchyard