GLM-5.3 Optimizes Its Own Inference Infrastructure

Zhipu disclosed that its GLM-5.3-Flash model's inference system, running on over 100,000 domestic AI accelerators, was optimized by an AI agent powered by GLM-5.3 itself. The agent achieved a 3.2x throughput improvement in under two weeks, marking a step toward recursive self-improvement.
Key insights
Core Achievement: Self-Optimization
Zhipu announced that its GLM-5.3-Flash model is now running production inference on over 100,000 domestic AI accelerators. The key milestone is that the optimization of the inference system was performed by an AI agent driven by GLM-5.3 itself, not by human engineers. The agent improved end-to-end throughput by 3.2x in under two weeks.
The 'Dense Feedback' Methodology
The agent's main challenge was diagnosing why performance degraded. Zhipu solved this by building a 'dense feedback' system with layered verification interfaces. This provides three types of signals: correctness (is the math right?), system behavior (where is time spent?), and performance (which strategy wins and under what conditions). This approach designs the environment for the agent rather than just giving it more capability.
Case Study 1: Numerical Precision Drift
The agent discovered that the KDA kernel's precision drifted with sequence length during context-parallel paths. TF32 rounding errors accumulated through chain state matrix merging. The agent identified the root cause and the fix was merged into the open-source project Flash Linear Attention.
Case Study 2: Concurrency Bottleneck
The agent found that KV cache transfer and DeepEP dispatch were never overlapping as intended. By tracing the call chain across the Python/C++ boundary, it discovered that the intra-node path in DeepEP was not releasing the Global Interpreter Lock (GIL), preventing the transfer thread from executing. The fix reduced overhead from over 30% to less than 1%.
Case Study 3: Redundant Computation
A decode kernel was performing the same normalization operation four times due to its blocking strategy. The agent restructured the computation, achieving a 1.71x speedup. The optimization idea came from 'optimization skeletons' the agent distilled from reading kernel code in projects like SGLang, FLA, and DeepGEMM.
The Recursive Self-Improvement Loop
Zhipu CEO Tang Jie emphasized that this creates a closed loop: the model optimizes the system, and the system serves the model. The layered verification environment built for the agent also serves as ideal training data for future models. Each completed task becomes a training example for the next generation, creating a data flywheel.
Transformation of the Engineer's Role
Tang Jie stated that the engineer's role is shifting from solving problems to designing feedback systems. Humans are still responsible for defining goals, building the feedback environment, and reviewing high-risk changes. The agent handles the execution and diagnosis within that structured framework.