GLM 5.3

Z.ai's flagship generation for coding and long-horizon agentic work, and an unusual release in that it shares GLM-5.2's base model entirely -- every gain comes from scaled-up post-training. Z.ai reports it as the strongest open-weights coding model it has measured, with roughly a 50 percent improvement over 5.2 on its own coding evaluation, and it reaches those scores while spending fewer output tokens per task. A capability the lab says grew faster than expected is vulnerability analysis: it reasons across multiple stages of exploitation rather than spotting isolated flaws. Thinking is mandatory, at three effort levels.

The unusual thing about GLM-5.3 is what did not change. It runs on GLM-5.2's base model, unmodified - Z.ai's own summary of the release is that scaling post-training is all it did. Everything below is therefore a claim about how much is still available after pretraining stops, which makes this release readable as an experiment as much as a product.

The coding gains are the headline. Z.ai calls it the strongest open-weights coding model it has measured, and the numbers it reports are large enough to be worth stating individually: Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Agents' Last Exam from 23.8 to 28.5. It attributes the durability of those gains to reinforcement-learning strategies carried over from 5.2, including compaction, which it says is what makes the improvements hold on long-horizon tasks rather than only on short ones.

The efficiency dimension is where Z.ai draws its sharpest comparison, because it measures score against tokens spent rather than score alone. On its in-house Z.ai Code Bench - a private benchmark it built to reduce contamination from public test sets - it reports 34.5% at roughly 75K output tokens per task at Max effort, against 23.4% at 96K for GLM-5.2: better and cheaper at once. At High effort it reports 31.4% at around 50K output tokens, ahead of Claude Opus 4.8's 29.5% at 120K. Z.ai is also explicit about where it stops: it says GLM-5.3 remains behind Claude Fable 5, which reaches 39.5% at Max effort.

The part Z.ai says surprised it is security. It added vulnerability-discovery data and environments to the post-training mix expecting the model to get better at spotting flaws, and reports that the capability kept developing faster than anticipated as training scaled - the model began reasoning across multiple stages of exploitation and forming coherent plans for complete chains, rather than identifying isolated bugs. Run against real codebases with security teams, and after expert review and deduplication, it reports 2,436 vulnerabilities found across 269 projects, 1,097 of them medium-to-high severity, spanning kernels, operating systems, browser engines, infrastructure, web applications and network protocols. Many had gone unnoticed for years; the oldest dated back around four decades.

Thinking is mandatory on this model and runs at three effort levels. Z.ai says the weights would follow the API release by about two weeks, after safety evaluation and hardening.