GLM-5.3: Open-Weight Coding Leader, Emergent Cyber Skills

GLM-5.3 is Z.ai’s new flagship coding model, and it arrives with two unusual claims: open-weight SOTA on the hardest agentic coding benchmarks, and “emergent” cybersecurity capability that developed faster than the lab expected. Released August 14, GLM-5.3 keeps the GLM-5.2 base model and gets every gain from post-training — no retraining of the base, no new architecture.

The headline numbers are the jumpiest the open-weight world has seen this year: Terminal-Bench 3.0 goes from 4.6 on GLM-5.2 to 28.3 on GLM-5.3 — an open-source SOTA — and CyberGym hits 84.5, ahead of every closed frontier model in Z.ai’s comparison table.

Key facts

• Released August 14, 2026 — post-training upgrade on the GLM-5.2 base; 1M-token context.
• Coding: Terminal-Bench 3.0 at 28.3 (open-source SOTA; GLM-5.2 scored 4.6), DeepSWE v1.1 at 66.9 (up from 46.2), Agents’ Last Exam at 28.5 (up from 23.8), AutomationBench at 48.2.
• Cyber: CyberGym 84.5 (state of the art, per Z.ai), ExploitBench 54.4 vs 24.4 on GLM-5.2, ExploitGym 105/130 vs 29/39 — more than doubling the previous generation.
• Weights in ~2 weeks — “We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.”
• Access today: GLM Coding Plan and ZCode; the API is “coming soon” per official docs.
• Z.ai Code Bench: a new in-house benchmark for coding agents in realistic developer environments; GLM-5.3 improves 50% over GLM-5.2 on it.

What changed

GLM-5.3 is a pure post-training story. Z.ai scaled RL with longer, richer agent environments and a verifier that passes oracle, no-op, and unsolved-state checks, then trained on the binary reward directly. The gains concentrate exactly where the industry’s benchmark pain points are: long-horizon terminal work, repo-scale engineering, and tool use — the Terminal-Bench family of evals we’ve been tracking since July.

The cyber results are the genuinely new thing. Z.ai says cyber capability “developed faster than we expected” as post-training scaled — GLM-5.3 is SOTA on CyberGym for vulnerability discovery and more than doubles GLM-5.2 on exploitation benchmarks. That capability is why the weights are delayed: safety evaluation and hardening come first, with open weights promised roughly two weeks after launch.

Why it matters

• Open weights keep closing the agent gap. GLM-5.3’s coding numbers sit within a few points of closed flagships on several evals, continuing the arc we charted in our best open-weight coding models ranking — and its DeepSWE score beats several paid APIs outright.
• The security dimension reframes “open.” A lab holding back weights for safety reasons while selling API access is a new stance in the open-weight world — a middle path between Qwen’s instant Apache-2.0 releases and closed labs. Whether the two-week window holds will be the test.
• Chinese labs now set the open benchmark pace. GLM-5.3, Qwen3.8-Max, and DeepSeek V4-Flash all shipped open-weights SOTA claims within six weeks; Z.ai’s 743B-class model is the latest proof that post-training, not pretraining scale, is the current frontier differentiator.

What to watch

• The two-week weights window. If the weights land with full evaluation notes, GLM-5.3 becomes the strongest open coding model on paper; if the window slips, the “safety evaluation” framing gets tested.
• Third-party benchmark runs. Z.ai’s table mixes vendor-reported figures; independent runs will confirm whether Terminal-Bench 3.0’s 28.3 and CyberGym’s 84.5 hold under other harnesses.
• The API launch — Z.ai says GLM-5.3 is coming to its API; pricing and availability will matter for the open-weight API pricing map.
• How closed labs respond to a permissively licensed model that outperforms them on vulnerability discovery — expect system-card language to get more careful.

Official source

• Z.ai: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities (August 14, 2026)
• Z.ai: Developer docs — GLM-5.3 overview

Updated August 16, 2026.