Inside GLM-5.3's Post-Training Factory: A Frozen Base, Environment Synthesis, and 1e-7 Alignment Carrying Every Gain

On August 14, Z.ai released GLM-5.3. The most unconventional thing about this launch is not the benchmark numbers but the opening line of the official blog — "Scaling post-training is all we did for GLM-5.3." The base model is identical to GLM-5.2: the same 743B MoE, not a single parameter changed, with every capability gain coming from the post-training phase.

Freezing the Base Is an Engineering Decision

Re-pretraining a hundred-billion-class model takes tens of thousands of GPUs, months of time, and massive compute investment; freezing the base means downstream inference infrastructure and operator optimizations migrate at zero cost, and the community's fine-tuning work on the previous generation remains fully compatible. As marginal returns on pretraining parameter scale narrow, the battleground shifts from "who has more parameters" to "who trains better in post-training."

Supporting this route is a three-part stack (the official blog links all three):

  • IndexShare (paper arXiv:2603.12201): efficient long-context processing;
  • SAO (paper arXiv:2607.07508): RL for long-horizon tasks;
  • slime (open-source repo THUDM/slime): a large-scale asynchronous training framework, with Megatron on the training side and SGLang on the rollout side.

The Real Bottleneck: Not the Model, the Environments

The most information-dense sentence in the official blog is this: as agent capability improves, much of the difficulty in scaling post-training moves from the model to the environment. A useful task environment has to be executable, verifiable, and close to real professional work — and you need many of them, not a handful of hand-built samples.

Z.ai's answer is an end-to-end environment synthesis pipeline:

  1. A Research Agent collects task patterns from real work and turns them into runnable long-horizon environments with multi-step dependencies and hidden state;
  2. A Judge Agent attempts each task to verify it is actually solvable;
  3. Verifiers are synthesized — note that the verifier has no access to the reference solution during synthesis;
  4. Solver trajectories are used to discover and close reward shortcuts;
  5. After passing oracle, no-op, and unsolved-state checks, the verifier produces a binary reward reliable enough to train on directly.

The pipeline still requires a meaningful amount of human-in-the-loop work; the team admits making environment generation and verification more autonomous is one of the next steps.

slime: The Hard Engineering Inside the Training System

All training runs on slime. Its design keeps training, rollout, and the data buffer on a single dataflow — math, code, sandboxes, verifiers, and long-horizon agentic environments plug in as data generation rather than as changes to the training loop. That is what let Z.ai keep adding environments from GLM-5.2 through 5.3 without rebuilding the training stack.

This iteration added top-p mask, top-k and full-vocabulary OPD on the algorithmic side, plus configurations improving training-rollout consistency. The hardest number: the average logprob difference between the training and rollout sides is controlled at the 1e-7 level, a reduction of more than 99.99% versus previous setups. On the systems side, local storage serves as an additional caching layer; multi-teacher OPD supports dynamic teacher switching with prefetching — no dedicated long-running inference service per teacher — and with joint scheduling and load balancing between the router and slime, the team reports end-to-end RL training throughput on long-horizon coding improved by more than 2.3x.

Results: Beyond Benchmarks, Read the Token Bill

On results: Terminal-Bench 3.0 rose from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and CyberGym from 77.2% to 84.5% (ahead of Mythos 5's 83.8% and GPT-5.6 Sol's 83.6%, per the official post). More telling is token efficiency on the in-house Z.ai Code Bench: at High effort, GLM-5.3 reaches 31.4% accuracy at roughly 50K output tokens, surpassing Claude Opus 4.8's 29.5% at 120K tokens — better results at less than half the cost.

There is real-world validation too: working with several security teams in China, after expert review, screening, and deduplication, the model identified 2,436 vulnerabilities across 269 projects (107 Critical, 990 High), with the oldest flaw dating back to 1981 and the average vulnerability having lived 26.6 years in the codebase — all tracked in the public ledger at cvd.z.ai. Weights will be released two weeks after launch, once safety evaluation and hardening are complete.

So What

The signal GLM-5.3 sends to the industry is direct: as base-model parameter scales converge, "whose environment factory is built better and whose RL system runs more stably" becomes the new dividing line. An environment synthesis pipeline + 1e-7-level training consistency + 2.3x throughput — less sexy than benchmark scores, yet they determine the cost and speed of the next capability jump. The next time a model "suddenly gets stronger," it is worth asking first: how were its environments built?

Sources: Z.ai official blog, HappyRock deep dive (2026-08-16)