Open-Source LLM Arena Shift: How Post-Training-Driven Intelligence Leaps Are Reshaping the Industry

A Notable “Tie”: The Technical Divide Behind the Numbers

A highly symbolic data point recently emerged in the open-source large model space: GLM-5.3 scored 60 on the Artificial Analysis Intelligence Index, tying with another closely watched open-source model, Kimi K3, for first place in the open-source camp, while sitting on the same level as several closed-source flagship models. If judged by total score alone, this would be just another routine headline about open source catching up to closed source. What is truly worth examining, however, is the difference in implementation paths behind that score—it reveals a profound paradigm shift in how large model capabilities are being improved.

An even more striking figure comes from the GDPval-AAv2 Elo rating. The model’s score jumped from 1524 to 1770, a single-iteration gain of 246 points, ranking second among all evaluated models—behind only the strongest closed-source model (1855 points) and more than 100 points above another open-source model at a comparable level (1668 points). A leap of this magnitude is exceedingly rare in the history of large model iterations, because it is not incremental optimization but a signal of something qualitative.

To understand the industry significance of this jump, one cannot look at the benchmark numbers alone. The scorecard needs to be read alongside deep reflections on “Scaling Law” published around the same time by core members of the model’s research team. Together, the two paint a complete picture of large model competition entering a new phase.

Two Fundamentally Different Technical Paths: Stacking Parameters vs. Squeezing Potential

In the current race among top models, Kimi K3 follows a classic “brute force” path: a total parameter scale of 2.8 trillion, a MoE (Mixture of Experts) architecture with 896 expert modules, of which 16 are activated per inference. This has been the mainstream approach of the large model arms race over the past two years—bigger base, more experts, higher training compute investment, trading physical “size” for capability “strength.”

GLM-5.3 chose a fundamentally different path: its total parameter count is 743 billion, and it shares the same base model as the previous version, GLM-5.2, with no changes whatsoever to architecture design or parameter scale. This means the entire performance gain came not from “heavy firepower” in the pretraining stage, but entirely from technical breakthroughs during post-training. Specifically, over the course of one month, the research team built long-horizon task training environments dozens of times larger than before, introduced a richer variety of environment types, and conducted larger-scale reinforcement learning training on that foundation.

One path frantically stacks parameters during pretraining; the other frantically squeezes the potential of an existing base during post-training. These two paths collided at the same point in time, forming one of the most noteworthy divides in the current large model competitive landscape. It tells us clearly: when base model scale growth hits a bottleneck or diminishing marginal returns, the post-training stage is becoming the decisive battlefield for final intelligence levels.

The Evolution of Scaling Law: More Than One Knob to Turn

To understand why post-training can deliver such a massive performance leap, it helps to review the historical trajectory of Scaling Law research. This theory is not an unchanging iron law but a dynamic process that is continually revised and reinterpreted.

Phase 1: The Era of Parameter Priority (around 2020)

Early classical Scaling Law research revealed an important pattern: model parameters should grow significantly faster than training data volume, at a ratio of roughly 2.7:1. This conclusion profoundly shaped the design philosophy of the first generation of ultra-large models, with the industry broadly following this guiding principle and pushing model scale to hundreds of billions or even trillions of parameters. The core logic at the time was: the larger the parameter count, the stronger the memory capacity and expressive capability, and performance naturally grows linearly with scale.

Phase 2: The Data Efficiency Correction (around 2022)

As research deepened, subsequent breakthrough work (represented by Chinchilla) corrected the early conclusion. New findings showed that allocating parameters and data at a 2.7:1 ratio is not optimal; the true golden ratio is roughly 20 training tokens per parameter. This correction pulled the industry’s focus directly from a “pure parameter race” back to “parameter-data balance,” prompting more teams to prioritize the collection and governance of high-quality training data.

Phase 3: Introducing Inference Cost (Current Phase)

However, the objective function optimized by Chinchilla still carries an implicit premise: a model is “trained once, evaluated once.” But in real production environments, a mature large model is invoked billions of times per day, and inference cost is the dominant factor in long-term operations. When inference cost is brought into the equation, the optimal solution changes—no longer “large parameters, short training,” but rather “small parameters, long training.” Under this approach, moderately controlling model size and investing more compute into training rounds and training quality can yield higher effective performance at the same or even lower inference cost.

Phase 4: New Insights Under MoE Architecture

For the currently mainstream MoE (Mixture of Experts) architecture, researchers have proposed a more illuminating framework: total parameter scale determines “how much the model can hold”—that is, the breadth of knowledge reserves, factual memory, and long-tail information coverage; while activated parameters and effective depth determine “how far the model can think”—that is, the length of reasoning chains, causal inference capability, and capacity for complex logic.

This distinction is critical. In practice, when the training compute budget (TPP, Total Processing Power) is fixed, blindly pushing up total parameter scale actually dilutes how thoroughly individual experts are trained, harming the model’s overall reasoning capability. Conversely, activating more experts on the same base reliably enhances the model’s reasoning depth. This finding provides clear theoretical guidance for post-training optimization: rather than starting from scratch to train a larger base, it is better to finely adjust activation strategies and reinforcement learning signals on an existing base.

Post-Training-Driven Intelligence Leap: Methodology and Engineering Practice

Against this theoretical backdrop, the technical evolution of the post-training stage has formed a relatively clear methodological system, with three core pillars.

1. Scaling Up Long-Horizon Task Environments

Traditional applications of reinforcement learning to language models have mostly focused on short-horizon, single-turn, clearly targeted interaction tasks. The breakthrough of the new generation of post-training techniques lies in building large numbers of task environments that require long-horizon planning, multi-step decision-making, and continuous trial-and-error. These environments simulate complex real-world problems—such as maintaining context across multi-turn conversations, modifying multiple files in programming tasks, and retrieving and synthesizing information in research tasks. Through trial and error in these environments, models gradually strengthen their capacity for “delayed gratification”—not pursuing single-step optimality, but choosing long-term strategies for the achievement of overall goals.

2. Diversifying Environment Types and Increasing Complexity

If scale is the guarantee of training intensity, then diversity is the key to training coverage. A single type of training environment easily leads to overfitting on specific tasks, whereas a rich variety of environment types—spanning different cognitive dimensions such as reasoning, creation, planning, and tool use—forces the model to abstract general problem-solving strategies. This “curriculum learning” style of arrangement lets the model start from simple tasks and gradually transition to high-difficulty, high-complexity tasks, thereby exhibiting stronger generalization on evaluation sets.

3. Fine-Grained Design of Reinforcement Learning Signals

The design of reward functions in post-training reinforcement learning directly determines the direction of model optimization. Current mainstream practice has upgraded from simple “outcome correctness” judgments to multi-dimensional process evaluation. For example, in code generation tasks, it is not enough to assess whether the final code passes tests; the logical soundness of intermediate steps must also be evaluated. In mathematical reasoning tasks, it must be assessed whether solution steps are clear and free of jump errors. This fine-grained signal design allows the model to more precisely locate its own weak points during exploration and achieve rapid iterative improvement.

Far-Reaching Impact on the Industry Landscape and Key Takeaways

The significance of GLM-5.3’s breakthrough far exceeds the rise or fall of a single benchmark score.

First, it proves that the post-training stage offers extremely high “cost-effectiveness.” Against a backdrop of tight compute resources and large model training costs running into tens of millions of dollars, being able to achieve substantial performance gains without changing the base model—purely through adjustments to reinforcement learning training strategy—is a highly attractive direction for many AI research teams. It lowers the barrier to entry for “frontier models”—no longer can only giants with tens of thousands of GPUs compete.

Second, it shakes the entrenched assumption of “parameter scale determinism.” The industry has long harbored a thinking inertia: bigger models must be stronger. But this result clearly shows that with a total parameter gap of nearly 4x, post-training methods can close and even reverse that gap. This suggests that when evaluating a model’s true capability, one should not look only at its “book value parameters” but also at the effectiveness of its training strategy and data utilization efficiency.

Finally, it injects a shot of confidence into the open-source community. Open-source models matching and in some cases partially surpassing closed-source flagships in evaluations demonstrates that the open collaboration technical path still has strong vitality. If post-training methods can be fully understood and replicated by the community, we have reason to expect more teams to achieve similar performance leaps on their own base models, thereby driving the prosperity of the entire large model ecosystem.

Future Outlook: The Focus of Competition Is Shifting

It is foreseeable that the next phase of large model competition will show two notable trends. First, the “arms race” in pretraining base models will not stop entirely, but its focus will shift from “absolute parameter scale” to “architectural design efficiency”—for example, how to achieve stronger reasoning depth with fewer parameters. Second, the post-training stage will become the primary field of differentiated competition, with teams engaging in fine-grained contests across dimensions such as reinforcement learning environment construction, reward model design, and alignment strategy optimization.

For AI technology decision-makers and developers, the core takeaway from this event is: when evaluating and selecting large models, one should not be misled by parameter numbers or surface-level benchmark scores, but should examine in depth the training methodology, the degree of post-training investment, and real reasoning performance across different scenarios. Future intelligence levels will increasingly depend on “how to train” rather than “how big to train.” This is both an inevitable outcome of technological evolution and a sign of the industry maturing.