Jiuzhang AI Cloud Unveils RL Infrastructure to Power Post-Training Scaling of GLM-5

Leiphone said that as pre-training yields diminishing returns, advanced capabilities like mathematical reasoning, code generation and complex agent tasks increasingly rely on RL. This requires not only better algorithms but also infrastructure that can efficiently manage continuous generation, feedback and iteration. Jiuzhang AI Cloud, in partnership with model developers, has deployed GLM-5 series and DeepSeek R series on its cloud for large-scale token production.

Jiuzhang AI Cloud's system consists of three core components: a generator that produces rollouts, an environment that executes tasks and returns rewards, and a trainer that updates model weights. The key challenge is matching trainer throughput with generator throughput to avoid idle computing or stale rollouts. For long-horizon agent tasks, a single rollout may involve multiple tool calls and long contexts, making the generator a full real-time inference system. Jiuzhang AI Cloud uses global dynamic scheduling to allocate resources flexibly across generation, training and environment phases. This approach, it says, boosted first-day speed by 1.5 times on models like MiniMax M2.1 229B and Qwen3-Coder-Next 80B, and a further 1.25 times with adaptive speculative training.

The company also emphasizes "state as a resource." It reorganized data flow with DingoFS distributed file storage, DFKV distributed cache, zero-copy data paths and RDMA networking, turning KV caches from temporary GPU memory into reusable state across tasks. This avoids repeated prefill and computation. Jiuzhang AI Cloud argues that inference efficiency directly affects training efficiency, as every high-quality token generation can be recycled as training data.

A joint empirical study with Renmin University's STILL project, released by Jiuzhang AI Cloud, verified that continuous RL iteration can elicit reasoning abilities in models like Qwen2.5-32B and DeepSeek-R1-Distill-Qwen-1.5B. The study also found that on-policy learning is key to sustained improvement and that a dynamic reward model method called Cooper can mitigate reward hacking.

Looking ahead, Jiuzhang AI Cloud said the same RL infrastructure can extend to embodied intelligence, where robots and agents interact with simulators and the physical world. The underlying needs remain consistent: elastic compute, real-time inference, state management, high-frequency feedback and continuous iteration.