GLM-5.3
Same base, post-training only
Z.ai announced GLM-5.3 on 14 Aug 2026: same base as GLM-5.2, every gain from post-training. Coding Plan is live that day. Standalone API and full weights are staged behind a safety review.
At a glance
Official post
z.ai/blog/glm-5.3. Opening line: Scaling post-training is all we did for GLM-5.3.
Versus 5.2
Officially the GLM-5.2 base — no new pretrain. Community still cites the ~744B 5.2 shape.
Who can use it
Coding Plan on day one. Standalone API is “coming soon.” Full weights in about two weeks, after safety work.
Thinking levels
Three effort levels in the integration note. OpenCode says Go has 5.3, 1M context, same price as 5.2 — that is OpenCode’s rate, not a Z.ai public tariff.
What 5.3 actually changed
Not a new base model. Z.ai says the jump came from more environments, more diverse tasks, and more compute on post-training. The post stresses harder coding, long-horizon work, and cybersecurity. Those cyber skills are dual-use, so release is staged: controlled partners first, broader API later, weights last.
What people are watching
OpenCode posted on 14 Aug: “GLM-5.3 now available in Go,” 1M context, same price as 5.2. Interconnects the same day: coding plan now, API soon, Hugging Face in two weeks. Vendor tables show jumps versus Kimi K3, Fable 5 and GPT-5.6 Sol. Independent reruns are not there yet — do not treat those rows as a ranking.
Timeline
GLM-5.2 ships and becomes the open-weight flagship of the line.
z.ai/blog announces 5.3. Coding Plan that day; OpenCode Go lists it the same day.
Official promise: publish full weights after safety evaluation. Standalone API still “coming soon.”
Confirmed vs treat carefully
Confirmed
Official blog dated 14 Aug 2026; same base as 5.2; Coding Plan first; API and weights staged; official account states cyber dual-use is why the broad release waits. Thinking levels low / high / max are in the integration note.
Treat carefully
Do not promote vendor Terminal-Bench / DeepSWE jumps as independent ranks. “Weights in two weeks” is not a download link today. Do not assume every reseller catalog added glm-5.3 on launch day.
Stay on 5.2 or chase 5.3
Keep GLM-5.2
You need MIT weights, a public API, and a Zhipu SKU you can order from a multi-model catalog today.
Wait / use a 5.3 channel
You already sit on Z.ai Coding Plan or OpenCode Go and can live with a staggered API and weights.
How to reach it
The official path is Z.ai Coding Plan. The standalone API id is previewed as glm-5.3; price and GA date are not locked on Z.ai’s public tariff. OpenCode users should read Go’s live model list. Local weights wait for a Hugging Face card — a promise is not a URL.
On QCode
As of this page, QCode’s /models list shows glm-5.1 and glm-5.2 only — not glm-5.3. This page does not claim the new SKU is callable here. For the Zhipu line today, use 5.2.
GLM-5.3 FAQ
Did 5.3 actually ship?
Yes. There is an official blog on 14 Aug 2026. That is not the same as every API or aggregator listing it.
Can I download the weights?
Not yet. Official wording: about two weeks, after safety evaluation and hardening.
What is different from 5.2?
Same base, no new pretrain. Officially the jump is post-training. Treat vendor benches as vendor benches until someone else reruns them.
Can I call 5.3 on QCode right now?
No. The catalog only lists glm-5.1 and glm-5.2.
What is the price?
Z.ai’s public price table still ends at 5.2. OpenCode’s “same as 5.2” is that channel’s rate, not a Z.ai tariff.
Does it have native vision?
The launch post sells coding, long-horizon agents and cyber. It does not pitch native vision as a 5.3 feature. Drop the pre-launch “5.3 will be multimodal” rumor.
Sources
z.ai/blog/glm-5.3 (14 Aug 2026); Z.ai’s official account on staged release and cyber dual-use; OpenCode, 14 Aug 2026, “GLM-5.3 now available in Go”; Interconnects, 14 Aug 2026.
Use the Zhipu SKU that is actually listed
GLM-5.2 is on the catalog. Do not treat 5.3 as in-stock here until /models says so.
Related
GLM-5.3 status tracker
The pre-launch page — “a string is not GA.”
GLM-5.2 guide
The Zhipu flagship you can actually call today.
DeepSeek V4 Pro 0813
The other open-weight flagship from the same week.
Not affiliated with Z.ai / Zhipu. Vendor benches are not re-run here. QCode stock is whatever /models shows live.