Jiuzhang AI Cloud Details Training-Inference Consistency in RL System with GLM-5 as Example
A Leiphone report outlines how Jiuzhang AI Cloud's reinforcement learning system achieves training-inference consistency, using GLM-5 to show efficient scaling of post-training RL.
The report says that as pre-training marginal returns diminish, large-model competition is shifting from sheer parameter stacking to post-training RL, which drives advanced abilities such as mathematical reasoning, code generation, complex decision-making, and long-horizon agent tasks. A complete RL system consists of a Generator, Environment, and Trainer, forming a closed loop of generation, reward, training, and update. The core challenge is maintaining dynamic matching between Trainer throughput and effective Generator throughput, as an imbalance can lead to idle compute or stale policy.
The report highlights GLM-5.2, developed by Zhipu AI, which scored 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 3.0, credited to RL training for long-sequence, multi-step and tool-use scenarios. GLM-5 series and DeepSeek R series are deployed on Jiuzhang AI Cloud for large-scale token production, with a collaborative model where model vendors and AI infrastructure jointly scale RL. The infrastructure treats context, KV Cache, and rollout data as state resources, using tools such as DingoFS distributed file system, DFKV distributed cache, zero-copy paths, and RDMA to reduce non-compute overhead and allow KV Cache reuse across tasks.
For models such as MiniMax M2.1 229B and Qwen3-Coder-Next 80B, the system achieved a 1.5 times speedup on the first day and an additional 1.25 times speedup over static predictors as traffic changed. Online speculative learning is implemented as asynchronous RL in production, with zero-downtime weight updates. The report cites an empirical study with Renmin University's STILL project showing that models like Qwen2.5-32B and DeepSeek-R1-Distill-Qwen-1.5B can improve through continuous RL, and that on-policy learning is important while reward hacking needs mitigation.
The report also references SemiAnalysis's RL system analysis, which states that only infrastructure with high alignment between training-inference throughput and system state can fully release RL potential. The report concludes that RL is evolving from a single training algorithm into a general infrastructure technology for large models, agents, and embodied intelligence, with Jiuzhang positioning its AI cloud as the core base connecting training, inference, state, and model iteration.