According to Z.ai CEO Jie Tang, parameter count alone no longer predicts AI capability; it must be considered alongside data quantity, compute allocation strategy, and deployment context. Instead, leading AI labs are discovering that how models spend their computing power during training and inference matters far more than their total size. This shift represents a fundamental rethinking of what makes an AI system actually intelligent.

Jie Tang recently outlined this new reality on social media, arguing that the traditional focus on model size misses the full picture. GLM 5.3, Z.ai's latest model, achieves significant performance jumps not through adding more parameters, but through reinforcement learning on complex, long-horizon tasks that mirror real professional work.

"Parameter count is only meaningful alongside three others: how much data you have, where you intend to spend your compute, and who will run the model, under what conditions," noted Jie Tang.

Jie Tang, CEO at Z.ai

What Changed in How AI Models Learn?

The traditional approach to building better AI models followed a simple formula: make them bigger. This logic, rooted in scaling laws established over the past decade, suggested that doubling parameters would yield predictable improvements. But that assumption breaks down when you look at what different types of tasks actually require. Memorization tasks, like recalling facts, do benefit from more parameters. But reasoning tasks, which require carrying long causal chains without losing the thread, respond better to different kinds of training.

GLM 5.3 achieved significant performance improvements partly through training on environments that simulate real engineering work, combined with reinforcement learning on long-horizon tasks. The model was given access to compute clusters, storage systems, internal documentation, and codebases, then asked to diagnose performance bottlenecks, implement optimizations, and deliver measurable improvements. Some of these tasks represent several days of work for an experienced engineer. This approach pushes models toward taking ownership of substantial work end-to-end, rather than relying on users to break problems into smaller steps.

The environments themselves are now being synthesized automatically. Research agents collect task patterns from real work and turn them into runnable scenarios with multi-step dependencies. Judge agents verify that tasks are actually solvable. Verifiers are created without access to the reference solution, and reward signals are discovered by analyzing successful solver trajectories. This creates a self-reinforcing loop where the training process itself becomes more sophisticated.

How Are Labs Rethinking the Scaling Equation?

Jie Tang identified multiple distinct factors of scaling that go beyond parameter count. These include data quantity, compute allocation strategy, model architecture choices like mixture-of-experts (MoE) sparsity, inference-time reasoning depth, and deployment context. The notation for MoE configurations has even evolved to capture this complexity more precisely.

The practical implication is striking: advanced skills like finding software vulnerabilities are not retrieval problems that can be solved by memorization. They require the model to maintain coherence across 20 or more inference steps without losing the thread of reasoning. Once a model reaches a certain knowledge threshold, this ability does not scale with total parameter count. Instead, it scales with the model's capacity to reason deeply during inference and its training on realistic, complex scenarios.

How to Evaluate AI Models Beyond Parameter Count

• Data Quality and Relevance: Models trained on synthetic, executable, verifiable environments that mirror real professional workflows show larger capability jumps than models trained on generic text data alone.
• Inference-Time Reasoning Depth: The ability to carry long causal chains during inference, rather than relying on memorized knowledge, is becoming the primary differentiator between capable and less capable models.
• Synthetic Environment Generation: Automatically synthesized task environments, verified by judge agents and refined through reward signal discovery, are replacing hand-built training datasets as the scaling frontier.
• Compute Allocation Strategy: Where a model spends its compute during training, on memorization versus reasoning tasks, matters more than the total amount of compute available.

This shift has immediate consequences for how AI labs allocate resources. Instead of racing to build the largest model, teams are now investing in better training environments, more sophisticated reward signals, and inference-time reasoning capabilities. The competitive advantage no longer goes to whoever has the most parameters, but to whoever can most effectively teach their model to reason through complex, multi-step problems.

The broader industry is already responding. Open-weight models like Ornith-1.5 are being released with end-to-end self-improvement capabilities, where models propose tasks, generate training scaffolds, and produce reinforcement learning rollouts to create new training experiences. These models are showing strong performance on agentic and coding workloads without necessarily being the largest in their class.

Meanwhile, smaller models are proving they can punch above their weight when trained on the right data. Microsoft researchers demonstrated that by learning from negative trajectories in coding tasks, a 32-billion-parameter model achieved a 4.1% improvement on a standard coding benchmark, reaching 58.05% accuracy on SWE-bench Verified. The key was teaching the model to recognize what went wrong in failed attempts, not just memorizing successful solutions. The approach involved extracting specific mistaken beliefs from failed trajectories and injecting them into training prompts as explicitly labeled wrong knowledge, so the model could learn what not to do.

The implications extend beyond raw capability. If reasoning and training strategy matter more than size, then frontier AI becomes more accessible to labs with smaller budgets but better ideas about how to structure training. It also means that the era of "bigger is better" is genuinely over. The next wave of AI progress will be driven by labs that can architect better learning environments, design more effective reward signals, and teach models to reason more effectively during inference, regardless of parameter count.