Jiuzhang Intelligent Computing Cloud demonstrates how reinforcement learning systems can enforce training-inference consistency for models like GLM-5, addressing diminishing returns from pure pre-training scaling.

As pre-training marginal returns diminish, model capability scaling is shifting from sheer pre-training accumulation to post-training reinforcement learning (RL). Advanced capabilities—such as mathematical reasoning, code generation, complex decision-making, and long-horizon agents—increasingly depend on RL to unlock their potential. By cycling through generation, execution, feedback, and reward, RL enables models to continuously refine their reasoning, planning, and self-correction abilities.

Consequently, large model development has transitioned from “one-time training” to “continuous training,” extending capability scaling across the entire generation-feedback-iteration lifecycle. SemiAnalysis notes that as RL becomes the cornerstone of continuous scaling, competition now hinges not only on algorithms but on the AI infrastructure systems that support continuous generation, training, and updates.

Zhipu has announced that GLM-5.3 and GLM-5.2 advance RL on the same foundational model, continuously raising the upper bound of model intelligence through post-training scaling. Taking GLM-5.2 as an example, it achieved a score of 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 3.0, driven primarily by RL training specifically designed for long-horizon, multi-step, and tool-coupling scenarios. This underscores a broader industry reality: advancements in complex reasoning and agent capabilities are increasingly reliant on reinforcement learning.

This paradigm shift is also reshaping the AI industry’s division of labor. Model vendors can no longer advance algorithms in isolation; they must collaborate with AI infrastructure providers to enable “continuous intelligent production.”

Currently, mainstream inference models—including the GLM-5 series and DeepSeek R series—are deployed on Jiuzhang Intelligent Computing Cloud for scaled token production. Moving beyond the traditional “model + compute” relationship, Jiuzhang and model vendors have established a collaborative framework for bidirectional RL scaling of algorithms and engineering systems. Vendors continuously push the boundaries of RL algorithms and strategies, while Jiuzhang refines adaptive compute scheduling, inference services, and state management for industrial deployment. As large-scale model algorithms evolve, they demand higher infrastructure standards, which in turn create the capacity for larger-scale RL implementation.

In this ecosystem, model vendors and AI infrastructure become two sides of the same coin. RL has evolved from a discrete algorithmic step in model training into a core capability that AI clouds must natively possess. Models, compute, data, state, and inference are no longer siloed; they jointly form a continuously operating intelligent production line.

The central question has therefore shifted from “model + compute” to how algorithms and infrastructure can co-scale to produce effective tokens and model capabilities at lower cost via “a single intelligent production line.”

The fundamental difference between RL and traditional pre-training lies not in the addition of a training algorithm, but in the restructuring of the training system itself.

A complete RL system typically comprises three components: Generator, Environment, and Trainer. The Generator receives prompts and generates rollouts via inference; the Environment executes code, tool calls, or task simulations and returns rewards; the Trainer updates model weights based on rollouts and rewards, then feeds the new weights back to the Generator. This forms a continuous loop: Generate → Reward → Train → Update → Generate. Unlike pre-training, which relies on static datasets, RL training data is generated in real-time by the model, marking the first true continuous production relationship between training and inference.

This architecture introduces a critical bottleneck. If the Trainer consumes 100 samples per second but the Generator only produces 60, compute resources sit idle. Conversely, if the Generator outpaces the Trainer, rollouts pile up and become stale, resulting in policy staleness.

Therefore, the core of RL infrastructure optimization is no longer isolated GPU utilization, but the dynamic matching of Trainer Throughput and Effective Generator Throughput.

Leveraging this mechanism, training and inference finally establish a genuine continuous production relationship. An RL system increasingly resembles a distributed production architecture: the Generator acts as the producer, the Trainer as the consumer, and the entire system functions as a distributed pipeline requiring continuous balance among throughput, data freshness, and resource utilization.

An empirical study jointly released by Jiuzhang Intelligent Computing Cloud and the STILL Project Team at Renmin University of China, titled *An Empirical Study on Eliciting and Improving R1-like Reasoning Models*, validates this mechanism. The research demonstrates that models such as Qwen2.5-32B and DeepSeek-R1-Distill-Qwen-1.5B can unlock latent reasoning capabilities through continuous RL iteration, with the latter achieving 39.33% accuracy on AIME 2024. The study further indicates that on-policy learning is crucial for sustained performance gains, while simply rewarding longer responses risks “length hacking.” The Cooper method, which dynamically updates the Reward Model, effectively mitigates reward hacking and improves end-to-end RL outcomes.

This proves that RL is far from simple incremental training; it is a highly coupled systems engineering effort involving strategy, generation, environment, training, and feedback.

Most easily overlooked is the dynamic matching between Generator and Trainer. If the Generator is too slow, training GPUs stall; if too fast, rollouts accumulate and age. For long-horizon agent tasks, a single rollout involves multiple tool calls and ultra-long contexts, making the Generator itself a complete real-time inference system. Furthermore, the two components share not just GPUs, but model weights, rollouts, KV caches, environment states, and task queues. Slow weight updates cause policy staleness, while overly frequent updates inflate model loading and data transfer costs.

Thus, RL infrastructure must solve not isolated increases in training or inference throughput, but maintain efficient matching between Generator and Trainer throughout the entire lifecycle.

Jiuzhang Intelligent Computing Cloud built its RL engineering infrastructure around this exact challenge: integrating Generator, Environment, and Trainer into a unified workflow, using global dynamic scheduling to adapt to resource demands at different stages. When generation becomes a bottleneck, inference resources are added; during training peaks, compute is reallocated; and when the Environment waits, idle GPUs are released. Consequently, AI scheduling shifts from “which machine has free GPUs” to “where should models, trajectories, states, and next-step computations reside.”

The direct benefit is accelerated model speed. On frontier models like MiniMax M2.1 229B and Qwen3-Coder-Next 80B, initial speed increased by 1.5x on day one; as traffic fluctuates, speed improved by another 1.25x compared to static predictors.

Meanwhile, most deployment schemes freeze offline-trained speculators, leading to slow deployment and failure under changing traffic or target models. Jiuzhang redefines online speculative decoding as an asynchronous RL problem running in production: accepted and rejected tokens serve as reward signals, training servers continuously update the speculator, and new weights can be hot-swapped, achieving zero downtime.

Contexts and KV caches from agent interactions, trajectory data and reward signals from RL generation, and real-time updated model weights collectively form the state assets for continuous AI evolution. If these resources are locked to a single GPU or node, cross-node transfers, repeated prefills, and redundant computation become new system bottlenecks.

Traditional data paths typically route through storage, CPU memory, framework buffers, and GPU VRAM, involving multiple copies and forwards. In continuous RL loops, these non-compute overheads accumulate and amplify. To address this industry pain point, Jiuzhang leverages underlying technologies like DingoFS distributed file storage, DFKV distributed caching, domain-wide zero-copy links, and RDMA high-speed networks to reorganize data and state flow.

Zero-copy shortens data transmission paths, reduces CPU involvement, and cuts non-compute overhead in RL closed loops. DFKV upgrades KV cache from a single-GPU temporary buffer to a locatable, migratable, and cross-task reusable state resource, allowing previously generated context to continue creating value.

Consequently, AI scheduling logic evolves from “where are free GPUs” to “where should models, compute, context, and states be combined.” If an agent has already performed extensive prefill on a node, subsequent requests can reuse existing KV caches, avoiding redundant computation. Weight exchanges between training and inference also leverage high-speed data paths to reduce transfer costs.

Based on this, Jiuzhang’s core philosophy for “training-inference consistency” is not merely sharing compute between training and inference, but enabling both to share a unified infrastructure capable of simultaneously sensing compute, data, state, and workflows. In RL frameworks, the Generator is inherently part of the training loop; in agent operations, context and KV caches are core states for continuous task execution, creating natural deep coupling. This means inference efficiency directly dictates training efficiency: every efficient token generation, KV cache reuse, and precise scheduling can translate into faster model iteration.

SemiAnalysis’s dedicated RL system analysis concludes that only infrastructure achieving high alignment between training/inference throughput and system state can fully unlock RL’s performance potential and support superior model capabilities. Jiuzhang, by focusing precisely on this, has achieved the first large-scale, full-scenario engineering validation in the Chinese market.

This “training-inference consistency” system capability serves as the foundation for RL, delivering significant advantages: accelerating inference optimization technologies to production-grade levels, which in turn boosts training efficiency and drives straight-line reductions in token costs.

Integration with MoE architectures, inference optimization, model quantization, model architecture design, and hardware co-design is particularly critical. For instance, prefill and decode exhibit distinct computational characteristics; PD separation matches these workloads to appropriate resource pools. Chunked prefill breaks down ultra-long contexts to alleviate resource competition across stages. Speculative decoding uses smaller models to generate candidate tokens and larger models to verify them in batches, reducing overhead from sequential token-by-token computation.

Though seemingly independent, these technologies essentially solve the same problem: producing more effective tokens from limited GPUs. Within Jiuzhang’s ecosystem, they converge into a single production line—the tokens generated by the Generator are not final answers, but raw materials for the next training round. Higher generator efficiency yields more rollouts per unit of compute; KV cache reuse eliminates redundant prefills; PD separation and dynamic scheduling further optimize compute-bandwidth matching.

Ultimately, training and inference optimization unify under a production-grade AI infrastructure objective: maximizing effective token output per unit of compute, and optimizing the conversion rate of tokens into model intelligence.

This represents the core value of Jiuzhang’s RL technology roadmap: rather than building an isolated RL training platform, it integrates generators, trainers, environments, model weights, KV caches, and GPU compute into a unified AI runtime. As model vendors intensify RL training, infrastructure simultaneously optimizes rollout generation, state reuse, and resource scheduling. Together, they determine how far post-training scaling can progress.

This capability will extend beyond large models to embodied AI. Agent tool calling, multi-turn decision-making, and robotic perception, reasoning, action, and feedback are fundamentally dynamic trial-and-error and continuous-learning RL loops. While future iteration environments will expand from code sandboxes and tool-calling scenarios to simulators, physical robots, and the real world, infrastructure requirements remain identical: elastic compute, real-time inference, state management, high-frequency feedback, and continuous iteration.

As the industry shifts from large-scale training to post-training reinforcement learning, RL is evolving from a single model training algorithm into a universal technological foundation connecting large models, agents, and embodied AI. RL enables continuous model capability scaling; Jiuzhang Intelligent Computing Cloud ensures this scaling occurs efficiently and at scale. This marks the pivotal transition of reinforcement learning from an algorithm to a core technology stack within AI cloud-native infrastructure.