XNews

Zhipu GLM-5.3-Flash optimized by its own AI agent

Zhipu GLM-5.3-Flash optimized by its own AI agent

Zhipu AI announced that its GLM-5.3-Flash model, 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 end-to-end throughput improvement in under two weeks, using a dense feedback system to identify and fix three critical engineering issues. The company stated this marks a step toward recursive self-improvement.

Key insights

AI agent optimizes its own inference infrastructure

Zhipu AI disclosed that the production inference for GLM-5.3-Flash is now fully running on over 100,000 domestic AI accelerators. The optimization was performed by an AI agent driven by GLM-5.3, not human engineers. The agent improved throughput by 3.2x from first run to production, within two weeks. CEO Tang Jie noted that the model is helping optimize the system that runs it.

Challenging deployment conditions on domestic accelerators

The deployment environment was constrained: limited memory and interconnect bandwidth on domestic accelerators, immature software stack with missing kernels and documentation. The model also required support for 1M token context and multimodal requests. Every optimization involved trade-offs: compute for memory (ReplaySSM), communication for memory (tensor parallelism), precision for capacity (mixed INT8/FP8/BF16 cache), and architecture separation for scheduling flexibility (encode-prefill-decode separation).

Dense feedback system enables agent to diagnose issues

The key insight was that the agent's bottleneck was not code writing but understanding why performance degraded. A 20% throughput drop provided sparse reward without root cause. Zhipu built a dense feedback system with three signal types: correctness (is the computation right?), system behavior (where is time spent?), and performance (which solution wins under what conditions). Each signal was designed to be local, cheap, and objectively verifiable, effectively giving the agent an environment to debug.

Three real engineering problems discovered and fixed by the agent

The agent identified and resolved three issues: (1) Precision drift in KDA context parallel path due to TF32 rounding errors accumulating in chain state matrix merges; (2) KV transfer and DeepEP dispatch not overlapping because the intra-node path did not release the GIL, causing overhead over 30% reduced to under 1%; (3) A decode kernel recomputing the same normalization four times due to block partitioning, accelerated 1.71x by restructuring computation. The fixes were merged into open-source projects.

Shift in engineer role from problem-solver to feedback designer

Zhipu emphasized that humans still define goals, build feedback environments, and review high-risk changes. However, the engineer's role is evolving from solving problems to designing feedback systems. The dense feedback environment serves dual purposes: it is both the agent's working environment and training data for future models, creating a data flywheel. CEO Tang Jie stated: 'We are far from recursive self-improvement, but the smallest loop already exists. The model optimizes the system; the system serves the model.'

Source posts