# The GLM-5.3 Setup Kit
GLM-5.3 is an open-weight coding model Z.ai released on 2026-08-14. On the company's own coding
benchmark it finishes 31.4% of tasks against Claude Opus 4.8's 29.5%, and it spends roughly 50,000
output tokens getting there where Opus burns 120,000. Agent runs are billed by the tokens a model
writes, so the second pair of numbers is the one that reaches your card.
The GLM-5.3 Setup Kit is a free three-file reference — plain markdown, no install, no account. It
covers plugging the model into a coding agent today, exactly what changes when the open weights
land, and which jobs to hand a cheap model versus which to keep on an expensive one.
## Start with the two things launch day is leaving out
It is not downloadable yet. Hugging Face says *coming soon*. Z.ai has indicated roughly two
weeks while safety hardening finishes, and as of today the only way to run GLM-5.3 is a paid coding
plan. Posts calling it free right now have not checked. That is why the kit is written in two
halves — a paid path you can start on this afternoon, and a free path that changes two environment
variables when the weights arrive.
It is not the best coding model. On the same chart that shows it beating Opus, Claude Fable 5
scores 39.5% and beats GLM-5.3 comfortably. What this release is, precisely, is the cheapest model
that stays in the frontier conversation. That is a genuinely useful thing to be. It is not the same
claim as winning, and the difference matters if you are about to move production work.
The numbers, including the ones that look bad
Terminal Bench 3.0 moved from 4.6 to 28.3 — open-weight state of the art, with the closed frontier
still ahead at 33.7 and 34.6. DeepSWE v1.1 went from 46.2 to 66.9. Every one of these gains comes
from post-training: the base model is the same one GLM-5.2 used.
Against Kimi K3, the other open model people are watching, GLM-5.3 clears terminal-agent work by
about ten points, 28.3 to 17.4. On DeepSWE, Kimi is narrowly *ahead* at 67.5 against 66.9. Both
facts are from the same comparison table, and reporting only the first one would be cherry-picking.
One caution that applies to all of it: these are the vendor's benchmarks, run on the vendor's
harness, with settings the vendor chose. That is not fraud, it is just weak evidence. The kit's
third file shows you how to run your own test on five real tasks from your own repository in about
twenty minutes, which is the only result that should move your setup.
The part that made its own makers wait
Z.ai added vulnerability-discovery training expecting better bug-spotting. What came out was
exploit-chain reasoning they had not planned for, and it is why the open weights were held back —
something that had not happened with a GLM release before.
On CyberGym, which measures finding real vulnerabilities in real code, GLM-5.3 scores 84.5%, ahead
of the closed frontier models at 83.8 and 83.6. That is first in the world on that benchmark.
It is not first everywhere in security: on ExploitBench it scores 54.4 against closed models at
78.0 and 76.5, so "best at security" is too broad a claim to make. But on the specific job of
finding vulnerabilities, a cheap model that leads the world beats an expensive model that is merely
good — and a false positive costs you five minutes of reading. Point it at your own repository.
Confirm anything it finds before you file it; a model's confidence is not a reproduction.
What to actually do
Route by the cost of being wrong, not by difficulty. Send boilerplate, scaffolding, mechanical
refactors, tests for existing code and first drafts you intend to review to the cheap model. Keep
architecture decisions, anything touching auth or money or deletion, and the final review before a
change ships on the expensive one. A cheap model producing a wrong test file costs a minute; a
cheap model producing a wrong migration costs a weekend.
Then hunt bugs in your own code, because that is the one job where this model is measurably first.
Why this kit survives the next release
The specific numbers here rot. The model id changes, the benchmark gets superseded, and in six
months something cheaper will ship. What does not rot is the shape: keep the model behind three
swappable values, and keep the choice of model separate from the code that uses it. Adopting the
next cheap model then costs an afternoon instead of a migration.
Comment GLM on the video and the kit comes to you directly.
*Sources: z.ai/blog/glm-5.3 · Hacker News item 49294997 (656 points, 319 comments on launch day).
Every figure above was checked against the release post and at least one independent write-up on
2026-08-14, including the ones that make the model look worse.*