ChinaModelAPI

News / Model Watch · Open-Weights Release

Open weights GLM-5.3 flagship 141 shards + BF16 custom glm-5.3 license
2026-08-29 · Model Watch · uploaded Aug 28 ~23:22 Beijing, on the promised window

GLM-5.3 Open Weights Land On Schedule: The Post-Training-Scaling Flagship Goes Downloadable

Zhipu closed the loop it opened on Aug 14: the GLM-5.3 flagship weights hit zai-org the night of Aug 28 — inside the promised ~two-week safety-review window, three days after the Flash sibling. The card's own headline: "Frontier Coding with Emergent Cyber Capabilities."

Direct answer

GLM-5.3 flagship weights are live: zai-org/GLM-5.3, 141 safetensors shards + BF16 variant, ungated, upload completed ~23:22 Beijing on Aug 28 (repo last-modified, API-verified). License is custom, named "glm-5.3" — read the LICENSE file before commercial self-hosting. Same base as GLM-5.2 with every gain from post-training (Terminal-Bench 3.0 4.6→28.3, AA Index 60, open co-#1); API already GA since Aug 19 at $1.40/$4.40.

What shipped (API-verified)

  • The repo. 153 files total, 141 safetensors shards, plus a separate GLM-5.3-BF16 repo; not gated. The HF page still renders an "Upcoming release" banner in some views — a page-template artifact; the files API confirms the artifacts are live and downloadable.
  • License. Custom, license_name "glm-5.3" (field: other) — not Apache 2.0. The Flash sibling ships its own terms; check both LICENSE files for the clauses that matter to you.
  • Timing, exactly as promised. "Weights in ~two weeks" (said Aug 14) → Flash early (Aug 26-27, via the Ox Alpha reveal) → flagship on the Aug-28 window. HN's "GLM-5.3 is now open-weight" thread: 537 points and climbing.
  • Positioning. Official card headline — "Frontier Coding with Emergent Cyber Capabilities" — leans into the defensive-security story (2,436 real vulnerabilities found, staged release for safety review) that defined this launch cycle.

Which GLM-5.3 weights should you pull?

Your workloadPullWhy
Max intelligence, coding + cyber + long-horizon agents (inference platform)GLM-5.3 (flagship)The post-training-scaling star; AA 60 open co-#1; datacenter-class serving
Multimodal (image/video in) at ~1/10 the costGLM-5.3-Flash320B/A18B, first natively multimodal GLM-5-series; local quantizations available
Hosted API, no infradocs.z.ai GA$1.40/$4.44 per 1M since Aug 19; glm-5.3 relay routing follows health-check

Flash specs per its official card (320B/18B active); flagship parameter class per launch-cycle reporting (~744B MoE) — confirm in the repo config before capacity planning.

The open-weight frontier, completed

With tonight's drop, every major Chinese lab's current flagship is downloadable within weeks of launch: Qwen3.8-Max (2.4T, plus Apache 27B and the Flash-Next architecture scout), Kimi K3 (2.8T — already post-trained into Harvey Tenet), and now GLM-5.3 + Flash. The frontier-vs-open gap that defined 2025 has, at least this quarter, closed from the Chinese side entirely.

Primary sources

FAQ (2026)

Weights really out?

Yes — 141 shards + BF16 on zai-org, upload completed ~23:22 Beijing Aug 28 (API-verified), ungated. HN announcement at 500+ points.

License?

Custom, named "glm-5.3" (not Apache). Read the LICENSE file before commercial self-hosting; Flash carries its own terms.

vs GLM-5.2?

Same base, all gains from post-training: Terminal-Bench 3.0 4.6→28.3, CyberGym 84.5%, AA 60 (open co-#1). Breaking change: thinking.type 'disabled' removed.

vs GLM-5.3-Flash?

Flagship = max intelligence (~744B-class, text-focused). Flash = 320B/A18B natively multimodal at ~1/10 price, local quantizations. Different jobs.

API instead?

GA since Aug 19 — $1.40/$4.40 per 1M on docs.z.ai; ZCode, Coding Plan, PhanRouter, JD Cloud, supercomputing platform all serve it.

What it completes?

The Chinese open-weight shelf: Qwen3.8-Max/27B/Flash-Next, Kimi K3, GLM-5.3+Flash — every current flagship downloadable within weeks of launch.

Related guides