XNews

GLM-5.3-Flash Inference Optimization by AI Agent

GLM-5.3-Flash Inference Optimization by AI Agent

Zhipu AI announced that GLM-5.3-Flash inference is now running on over 100,000 domestic accelerators, achieving a 3.2x throughput improvement via a GLM-5.3-driven AI agent that performed optimizations like ReplaySSM, tensor parallelism, and mixed-precision caching. Separately, Tencent open-sourced WeKnora, an enterprise knowledge platform integrating RAG, ReAct agents, and automated Wiki generation.

Key insights

GLM-5.3-Flash inference optimized by AI agent on domestic accelerators

Zhipu AI's GLM-5.3-Flash model now runs production inference on over 100,000 domestic AI accelerators. The entire optimization pipeline, achieving 3.2x end-to-end throughput improvement, was performed by an AI agent powered by GLM-5.3 itself. The agent applied techniques including ReplaySSM (compute for memory), intra-node tensor parallelism (communication for memory), mixed INT8/FP8/BF16 caching (precision for capacity), and encode-prefill-decode separation. The per-token cost on domestic chips matched leading NVIDIA GPUs.

Dense feedback mechanism enabled agent to diagnose three critical issues

The AI agent relied on a layered verification interface providing dense feedback on correctness, system behavior, and performance. This allowed it to discover and fix three real engineering problems: (1) precision drift in KDA kernel due to TF32 rounding error accumulation; (2) GIL contention preventing overlap of KV transfer and DeepEP dispatch; (3) redundant normalization computation in a decode kernel accelerated 1.71x. These fixes were contributed to open-source projects.

Shift from problem-solver to feedback designer for engineers

Zhipu CEO Tang Jie stated that the most important experience is not the optimizations themselves but the change in engineer role. Engineers now define goals, build feedback environments, and review high-risk changes, while the agent executes the optimization loop. This layered, verifiable feedback environment also doubles as training data for future models, creating a minimal recursive self-improvement loop.

WeKnora: enterprise knowledge platform with RAG, ReAct, and Wiki modes

Tencent WeChat team open-sourced WeKnora, an enterprise knowledge platform that transforms scattered documents into three assets: RAG knowledge bases for query answering, ReAct agents for multi-step reasoning, and automatically distilled Wiki knowledge bases with visualization and versioning. It supports multiple retrieval strategies (BM25, dense, GraphRAG) and comes with built-in evaluation tools for retrieval quality.

WeKnora enterprise features: multi-tenancy, security, and knowledge engineering

WeKnora includes multi-tenant support with four-level RBAC (Owner, Admin, Contributor, Viewer), AES-256-GCM encryption for credentials, TLS for gRPC/Redis, SSRF hardening, and sandboxed execution. Knowledge bases are treated like code repositories with versioned changes, auto indexing, and rollback. The platform is the core technology behind WeChat's dialogue open platform, providing production validation.

Source posts