In August 2026, competition among large language models showed a notable shift. Zhipu AI’s GLM‑5.3 and DeepSeek’s V4 series demonstrate in practice that once a model base reaches a certain scale, endlessly stacking pretraining data and parameters is not the only — or even the most economical — path to improving "usable intelligence." High‑quality post‑training, reinforcement learning (RL) environment design, and test‑time compute optimization can "squeeze out" substantial latent capabilities from an already strong base model.
This change is reshaping the industry's technical roadmap and cost structure.
1. Shared Signals from GLM‑5.3 and DeepSeek‑V4
GLM‑5.3 was officially released on August 14, 2026. Zhipu has stated clearly that it uses the exact same base model as its predecessor GLM‑5.2 (an MoE architecture on the order of ~740 billion parameters), and that all performance gains come from post‑training scaling: longer, richer task environments and much longer reinforcement learning training. Official reports indicate about a 50% improvement in internal programming‑skill evaluations versus the prior generation; Terminal‑Bench 3.0 rose from 4.6 to 28.3; DeepSWE v1.1 increased from 46.2 to 66.9; and the model claimed leading positions among open models on multiple agent and coding benchmarks. On CyberGym it reached 84.5%, approaching or slightly exceeding some closed‑source top models. Note that these figures are primarily vendor‑reported and independent third‑party reproduction is still ongoing.
DeepSeek‑V4 is even clearer in its public disclosures. Its preview release did not initially stand out on many benchmarks, but after a period of intensive post‑training its metrics improved substantially. The technical report describes a full pipeline of "domain‑expert RL → multi‑teacher On‑Policy Distillation (OPD) → unified model." The company first trains expert models for math, coding, agents, instructions, and other domains, then fuses those abilities into a single unified model via online policy distillation. The post‑training stage accounts for a relatively small share of total GPU budget yet yields marked gains in agent capabilities and long‑horizon tasks.
Both point to the same conclusion: when the base model is already strong, the emphasis shifts from "making the model know more" to "making the model better at using what it already knows."
2. Why can post‑training yield such large improvements?
Pretraining’s core objective is to teach the model "what the world is" — language, knowledge, patterns and latent regularities. Post‑training focuses more on "how to use those abilities when faced with concrete tasks."
A powerful base model may already possess knowledge of Python, algorithms, mathematics, Linux, databases and software engineering, but may not reliably close loops such as: understand the task → make a plan → call tools → observe results → detect errors → modify the plan → continue execution → verify the final result. These behaviors are essentially policy optimizations rather than injections of new factual knowledge. RL and high‑quality post‑training are particularly effective at tuning such policies.
Crucially, the design of the training environment matters. Traditional RL often relies on a reward model to provide scores, whereas recent advanced coding agent training increasingly uses "verifiable rewards": the model modifies code, which is executed with compilation and unit tests; passing tests yields reward. The reward signal comes from real execution outcomes rather than subjective human scores, substantially increasing reliability. Environment‑driven RL of this kind teaches models to actually perform tasks, not merely generate plausible answers.
3. DeepSeek‑V4’s refined path: expert isolation and online distillation
DeepSeek‑V4 further addresses multi‑capability conflicts. Different domains have conflicting reward objectives: math tasks prefer deep reasoning, casual conversation favors brevity, coding benefits from frequent tool use, and safety demands declining certain requests. Mixing all rewards naively leads to gradient interference.
DeepSeek therefore adopts a divide‑and‑conquer approach: it first brings domain experts to high performance through SFT and RL, then distills those capabilities back into a unified model via multi‑teacher On‑Policy Distillation (OPD). The key of OPD is "online" — the student model generates its own trajectories while teacher models provide full probabilistic guidance for the student’s current states, rather than relying on static ground‑truth answers. The student is corrected on its own mistakes, yielding learning efficiency far above traditional offline distillation.
DeepSeek also introduces a Generative Reward Model (GRM) to make evaluations more granular, enabling structured assessment of complex trajectories rather than a single scalar score. This further raises the quality and scalability of post‑training.
4. The real composition of capability gains: a three‑term product
State‑of‑the‑art effective intelligence can be approximated as:
effective intelligence ≈ base capability × policy quality × test‑time compute
- Base capability comes from pretraining and sets the knowledge and latent ceiling.
- Policy quality comes from post‑training and RL and determines how the model searches, acts, corrects, and verifies.
- Test‑time compute determines how much compute is actually invested during inference.
Thus, the same base can achieve large benchmark jumps through enhanced post‑training and allowance for higher inference cost, without repeating costly large‑scale pretraining. GLM‑5.3 exemplifies this logic: it does not change the base yet releases substantial latent capability through environment and policy optimization.
5. Risks to watch: leaderboard gaming, contamination and reward hacking
The power of post‑training also brings new evaluation challenges.
First is benchmark overfitting and data contamination. If the training environment closely matches public benchmarks, score gains may reflect "teaching to the test" rather than general capability. Second is reward hacking: models may learn to bypass test cases instead of genuinely solving problems. This risk is greater in agent settings, where models can exploit specific shortcuts rather than acquire broadly generalizable strategies.
Differences in test‑time compute are often overlooked. If a new model consumes several times more tokens on average during inference, a significant portion of benchmark gains may come from "allowing it to think longer" rather than from parameter or knowledge improvements. Accurate evaluation must disentangle these factors.
Therefore, figures like GLM‑5.3’s "50% improvement" should be interpreted cautiously: they likely mix genuine policy and agent capability gains with environment specificity, inference cost, and evaluation scaffolding contributions. Independent reproduction and performance on real long‑horizon tasks are the keys to judging true value.
6. Industrial implications: the competitive focus is shifting
GLM‑5.3 and DeepSeek‑V4 together indicate that frontier competition in 2026 has shifted from "who owns the stronger base model" to "who can most effectively train a base model into a policy that acts over time, uses tools, self‑corrects and verifies results."
Pretraining remains important because it sets the capability ceiling; but post‑training is becoming the key lever for unlocking usable intelligence and improving real task success rates. The value of verifiable environments and long‑trajectory RL has been convincingly demonstrated in coding, software use, and multi‑step agent scenarios.
The cost dynamics are also notable. Large‑scale pretraining consumes massive compute, while post‑training often takes a much smaller share of the overall budget yet can deliver substantial task performance improvements. For teams with limited resources, this provides a pragmatic path to catch up.
Conclusion: redefined capability boundaries
An already strong base model can, through carefully designed post‑training environments, verifiable rewards, online distillation and appropriate test‑time compute, release capabilities far beyond what traditional SFT achieves. This is not magic: the model already holds abundant latent knowledge, and post‑training changes how it searches, decides and corrects.
At the same time, the industry must adopt stricter evaluation standards: private benchmarks, fresh tasks, adversarial testing, real long‑horizon scenarios, and transparent reporting of inference costs. Only then can we separate "true capability gains" from "leaderboard illusions."
The post‑training paradigm represented by GLM‑5.3 and DeepSeek‑V4 is pushing large models from "knowledge containers" toward "action agents." This may be the deepest and most important change in current model competition, and it merits continued close attention.
▶️ Enjoy this piece? Producing it takes real effort!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are the greatest support! 🙏
More AI news at: https://www.airosetta.com