Chinese AI labs deploy recursive self-improvement without altering model weights

Zhipu, DeepSeek, ModelBest, Moonshot AI, Alibaba, ByteDance and Xiaomi are all reporting recursive self-improvement (RSI) work that targets training pipelines, inference systems and tooling rather than model weights, according to a Pandaily report on October 31, 2026 source.

Zhipu’s production example involves an Infra Agent driven by GLM-5.3 that built the serving stack for GLM-5.3-Flash on a cluster of about 100,000 China-made accelerators in two weeks, storing fixes as reusable templates; its HarnessPro paper with Tsinghua describes agents rewriting their own harness. DeepSeek’s DSec cluster, co-built with Tsinghua, produces about 3 million sandboxes per day, allowing agents to construct training environments, while its DeepEvolve agent separates search from forecasting. ModelBest’s Forge Engineering project claims its AI-written training framework ForgeTrain matched Megatron-LM performance in eight hours, and ForgeKernel operators topped Huawei and NVIDIA leaderboards.

Moonshot AI used over 51 million sandboxes on Firecracker microVMs for Kimi K3 training, employing an in-cluster agent to recover failed runs from 26 playbooks; a glitch in K3.1 Pro forced a reset of 25 percent of members’ monthly quota. Alibaba reports Qwen3.8-Max ran its own training pipeline for a month across 33 iterations, while ByteDance’s Seed team highlights a self-evolving model ID, and Xiaomi livestreamed six days of reinforcement learning on MiMo-V2.6.

Our read: The focus on RSI at the infrastructure level reflects a pragmatic shift in Chinese AI development, aiming to accelerate iteration cycles and reduce human engineering overhead without the regulatory and safety complexities of weight-level self-modification. The reported scale—hundreds of thousands of accelerators, millions of daily sandboxes and multi-week autonomous training loops—suggests that these labs are building reusable, agent-driven toolchains that could lower entry barriers for downstream B2B users, while also exposing new failure modes in automated pipeline control.