Tang Jie Discloses Zhipu's First RSI Case: GLM-5.3 Agent Builds and Tunes Its Own Inference Stack

The system carries all online inference traffic for GLM-5.3-Flash. According to the post, no one had previously deployed a domestic-chip cluster at that scale, and the team faced limited chip memory capacity and bandwidth, an immature operator and software ecosystem, incomplete documentation, and the need to support 1M-token context and multimodal requests on a new model architecture. Tang wrote that the job was completed not by a human team but by the GLM-5.3-driven Infra Agent.

The agent worked through operator precision problems, Python and C++ cross-layer concurrency bottlenecks, and kernel performance tuning, reading system feedback, forming hypotheses, editing code, running experiments and iterating on results. It traced a concurrency blockage caused by the Python GIL in a KV Transfer scenario, reducing a performance loss of more than 20 percent when combining Prefill with KV Transfer, compared with Prefill alone, to under 1 percent. On the KDA Decode operator, it reorganized the computation to obtain a 1.71 times speedup.

GLM-5.3-Flash itself was released anonymously as "Ox-Alpha" on OpenCode and OpenRouter for testing against real traffic, the post says. Within one week it became the most-called model on both platforms, and it recorded more than 62 trillion tokens of calls in six days.

The resulting stack combines intra-node tensor parallelism for linear attention and the LM Head, ReplaySSM, W8A8 quantization, INT8, FP8 and BF16 mixed-precision cache quantization, and Layer Split. An Encode-Prefill-Decode disaggregated architecture produced roughly a threefold improvement in end-to-end service performance, with hardware utilization and per-token cost reaching levels comparable to mainstream NVIDIA GPUs. The post puts the final gain at 3.2 times the initial baseline in one passage and about three times in another, and says the agent-assisted feedback loop compressed model adaptation to production readiness into under two weeks.

Much of the post is devoted to what Zhipu calls dense feedback. The company argues that code-level context alone is insufficient, because accuracy anomalies and performance regressions in an inference system emerge from interactions among operator implementations, parallelism strategies, communication behavior, memory management and service scheduling, while an end-to-end metric only reports that results got worse, not why. Zhipu folded correctness tests, runtime logs, execution traces, runtime events, microbenchmarks and end-to-end metrics into the agent's iteration loop, and describes feedback as useful only when it is local enough to narrow the problem, cheap and fast enough to obtain, and objectively verifiable.

One documented case involved a precision problem on the KDA operator's context parallel path. A deviation between CP and non-CP results pointed to state propagation and merging introduced by parallel execution. In the original implementation, tl.dot defaulted to TF32 even with FP32 inputs, and the lower precision accumulated error as state transformations were merged and updated, becoming more pronounced over long contexts. The fix explicitly set input_precision to "tf32x3", combining three TF32 tensor core operations to recover higher precision while retaining much of the performance benefit.

Tang wrote that the arrival of RSI remains distant. Goals, boundaries and risk judgments are still set by human engineers, and Zhipu states explicitly that it has not reached recursive self-improvement. As background, the post notes that the company began work on strengthening safety capabilities in October 2025 on the assumption that safety is a natural extension of coding ability, and that within a year security partners used GLM to find thousands of vulnerabilities in real codebases, prompting a trusted access program. Tang also wrote that before GLM-4.7, internal use of GLM for coding was partly obligatory within the company, whereas GLM-5.3 has become a daily coding tool.

He summarized the emerging loop in one line: the model optimizes the system, and the system carries the model.