Zhipu today published the first engineering RSI (Recursive Self-Improvement) case from a Chinese LLM vendor on its official technical blog. A GLM-5.3-powered Infra Agent designed, debugged, and optimized the entire production-grade inference stack for GLM-5.3-Flash from scratch, deployed on a cluster of more than 100,000 Chinese-made AI accelerators. End-to-end throughput tripled relative to the initial baseline in under two weeks, while hardware utilization and per-token cost reached levels comparable to mainstream NVIDIA GPUs.
RSI in name refers to systems that iteratively improve themselves and then help train their own successors. That endpoint still looks far off, but the early form is now running in production.
Getting a new model from its first successful run on unfamiliar hardware to a high-throughput inference service that can reliably handle production traffic has historically taken a team of infrastructure engineers weeks or months. Zhipu stacked three complications this time: limited chip memory capacity and bandwidth, a 1M-token context window, multimodal traffic, an immature ecosystem, and incomplete kernel support. The Infra Agent carried most of the work that used to live with the systems team.
The core idea: dense feedback that engineers can still sign off on
The team's central contribution is what they call "dense feedback," a feedback discipline that ties every signal back to a specific engine launch parameter, code change, kernel, input condition, or code path, rather than dumping more logs and metrics onto the agent. Three kinds of feedback answer three kinds of decisions: correctness feedback tells the agent whether the computation is right; system-behavior feedback pinpoints where time is going; performance feedback decides which approach actually wins under the target workload. The three layers convert the diagnostic chain that used to live inside an experienced engineer's head into a workflow an agent can execute continuously.
The blog walks through three concrete cases.
Case 1: numerical accuracy in a KDA kernel
The agent compared partitioned vs. unpartitioned kernel outputs and uncovered an accumulating-error bug on the KDA kernel's Context Parallelism path. The root cause was tl.dot defaulting to TF32 even with FP32 inputs, which compounded during state merging and updates. Setting input_precision="tf32x3" pulled the error back within tolerance. The fix was merged upstream into Flash Linear Attention as PR #1180.
Case 2: a missing GIL release that blocked KV Transfer
The performance gap between Prefill + KV Transfer and Prefill alone exceeded 20%, well past the 5% acceptance bar. Tracing the execution timeline showed that the Python Mooncake Transfer thread in the same process could not acquire the GIL promptly because DeepEP v1.2.1's intranode_dispatch and intranode_combine never released it explicitly. Releasing the GIL during the relevant C++ intervals closed the gap to under 1%.
Case 3: collapsing redundant V-dim tiles for a 1.71x kernel speedup
The agent extracted optimization techniques from existing kernels across SGLang, Flash Linear Attention, and DeepGEMM, then turned them into "optimization skeletons" with applicability conditions, transformation methods, and validation evidence. On a representative KDA Decode kernel, after ReplaySSM and a 9.6% reduction from division optimization, the agent noticed the original implementation tiled along the V dimension, repeating FP32 normalization and gating four times. Merging those tiles into a single thread block and replacing per-tile reductions with a single warp-level reduction yielded a 1.71x speedup over the v2 baseline.
The real signal
Zhipu is clear that choosing objectives, setting boundaries, and assessing production risk remain human responsibilities. But two weeks, threefold throughput, and 100,000 accelerators together say one thing: once the loop of "model optimizes system, system runs model" works in production, the engineering headcount required to ship the next frontier model starts to drop fast. That is the RSI story actually worth watching.