Chinese AI labs aim self-improvement at training infrastructure, not model weights, Leiphone survey finds

Chinese model developers are pursuing recursive self-improvement, meaning AI that helps build the next AI, by applying it to training and serving infrastructure rather than to model weights. A Leiphone survey of recent disclosures, reported by Pandaily, finds that no lab is letting AI rewrite weights. Labs instead target training pipelines, inference systems, GPU kernels, sandboxes and agent harnesses.

A framework for the trend

A September paper, The Last AI Built by Humans, by 35 researchers from Shanghai Jiao Tong University, Tsinghua, ByteDance, SenseTime and others, proposes five levels of self-improvement. It places most frontier systems between levels one and two.

What each lab disclosed

Zhipu has the clearest production case. An Infra Agent driven by GLM-5.3 built the serving stack for GLM-5.3-Flash in two weeks, on a cluster of about 100,000 China-made accelerators. The agent stores working fixes as reusable templates. Zhipu's HarnessPro paper, written with Tsinghua, then had agents rewrite their own harness.

DeepSeek started with the training ground. Its DSec cluster, built with Tsinghua, produces about 3 million sandboxes a day, so agents can build environments that train stronger agents. DeepSeek Harness treats tools and even its core runtime as plug-ins that can be swapped without a restart and rolled back. An October agent, DeepEvolve, separates searching from forecasting.

ModelBest runs a Forge Engineering project in which AI writes the software used to build AI. The company says its AI-written training framework, ForgeTrain, matched Megatron-LM in eight hours, and that its ForgeKernel operators topped official Huawei and NVIDIA leaderboards.

Moonshot AI relies on large-scale reinforcement learning. Training Kimi K3 used more than 51 million sandboxes on Firecracker microVMs. It also used an in-cluster agent that recovers failed runs from 26 playbooks, and an agent reward model that scores training recipes. The risks appeared in October, when sub-agent calls in K3.1 Pro spiraled and Kimi reset 25 percent of members' monthly quota.

Alibaba says Qwen3.8-Max ran its own training pipeline for over a month across 33 iterations. Qwen is also the base for outside efforts such as Apex-35B-V1.0.

ByteDance's Seed team offers a self-evolving model ID. It argues that most systems are only half-closed loops, because humans still set the scoring rules.

Xiaomi livestreamed six days of reinforcement learning on MiMo-V2.6.

The unresolved problem

Zhipu chief scientist Tang Jie says the engineering loops are real, but the hard part remains getting a model to judge reliably when to stop, when to correct itself and how to evaluate its own training.

What this means

The disclosures show self-improvement working where outcomes can be checked: a serving stack that runs, a kernel that tops a leaderboard, a training run that recovers from failure. Weights are left alone, and the survey frames the approach as engineering automation rather than models modifying themselves.

The Kimi K3.1 Pro quota reset shows that agent-driven systems can fail at scale in production. Seed's point about half-closed loops and Tang Jie's comments both place the remaining difficulty in evaluation: humans still define the scoring, and models cannot yet reliably judge their own training. Several of the performance claims, including ModelBest's benchmark and leaderboard results and Alibaba's account of Qwen3.8-Max, come from the companies themselves.