GLM-5.3 Arrives for Coding and Cyber Work, with Weights Still Pending

Z.ai announced GLM-5.3 at 05:17:49 UTC on 14 August 2026, or 07:17:49 SAST. The company made the model available immediately through GLM Coding Plan and ZCode. Its official follow-up says maker-hosted API access and open weights will be released in stages after safety evaluation.

The release uses the same 743-billion-parameter base model as GLM-5.2. Z.ai says the improvement comes from further post-training rather than a new base model. GLM-5.3 supports low, high and max thinking effort. Non-thinking mode is no longer supported.

Z.ai reports that GLM-5.3 rises from 4.6 to 28.3 on Terminal Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 23.8 to 28.5 on Agents' Last Exam. It also reports better results than GLM-5.2 on its private Z.ai Code Bench while using fewer output tokens at comparable effort levels.

Those numbers come from Z.ai and should not be treated as independent validation. The private benchmark may be useful for internal comparison, but its task set and scoring cannot be inspected in the same way as a public benchmark. Teams should test the model on their own repositories before changing production routes.

Z.ai says GLM-5.3 reaches 84.5 on CyberGym, 54.4 on ExploitBench, and completes 105 ExploitGym tasks within a two-hour normalised budget. The reported gains over GLM-5.2 grow as tasks move from vulnerability discovery towards exploitation.

The caveat matters. Z.ai's own chart shows closed models still ahead on ExploitBench and ExploitGym. The company is also delaying weights for two weeks while it completes safety evaluation and hardening. That delay is part of the release, not a minor footnote.

AIMI first observed opencode-go/glm-5.3 at 05:59:36 UTC, exactly 41 minutes and 47 seconds after Z.ai's announcement. It first observed cline/cline-pass/glm-5.3 at 06:37:42 UTC, 38 minutes and 6 seconds after the OpenCode Go sighting and 1 hour, 19 minutes and 53 seconds after the maker announcement.

These timestamps describe provider route availability, not separate model releases. Cline's catalogue description calls GLM-5.3 an open-weights model, but Z.ai says the weights are still pending. The maker's statement takes precedence. The endpoint times are observations from the AZ Labs monitor, not provider publication timestamps.

Dots Studio announced dots3-note preview at 04:00:31 SAST. GLM-5.3 followed at 07:17:49 SAST, exactly 3 hours, 17 minutes and 18 seconds later. That makes GLM-5.3 the second verified maker release in the 14 August AIMI input with a precise announcement time.

DeepSeek V4 Pro 0813 was also active in today's provider timeline, but its maker release is dated 13 August. Meta Muse Spark 1.2 was announced on 5 August. Their 14 August sightings were route expansion, not additional maker releases. No primary source links the timing of GLM-5.3 to Dots Studio or DeepSeek, so no competitive response is claimed.

The official Cline catalogue now lists cline-pass/glm-5.3 as part of ClinePass. AZ Labs sells ClinePass access, so customers can use the order link near the top of this article to request the plan. Availability and delivery are confirmed in the ticket before payment.

A ClinePass listing proves access through that provider. It does not prove that Z.ai's own API is generally available, that the weights have shipped, or that every capability shown in the maker's evaluation is exposed through the Cline route.