PublishersGLM-5 Team

GLM-5 Team

1 document read from this publisher.

GLM-5: from Vibe Coding to Agentic Engineering

research paper · technical report · 2026-02-17 · 53 techniques

Mixture of Experts Multi-Token Prediction DeepSeek Sparse Attention Group Relative Policy Optimization Expert Parallelism Gated DeltaNet Muon Sliding Window Attention Asynchronous reinforcement learning Discard-all context management Cosine decay Prefix caching Agentic data synthesis pipeline Attention data parallelism Batch-size warmup Prefill-decode disaggregation Asynchronous scheduling IcePop Lightning Indexer Multi-Head Latent Attention Token-in-token-out (TITO)Agent-as-a-Judge Asynchronous reinforcement learning infrastructure Direct Double-Sided Importance Sampling Multi-query attention On-Policy Cross-Stage Distillation RadixCache Rejection sampling Asynchronous Agent RL algorithms DP-aware routing FlashComm Flex_AWQ_SSZ FP8 inference Group-Wise Policy Optimization Heartbeat-driven fault tolerance Hierarchical context management Hybrid reward system INT4 Quantization-Aware Training Keep-recent-k masking-based refinement