GLM-5.3
Zhipu AI · Undisclosed (MoE, full Ascend stack)
An S-tier model trained end to end on Ascend
A minor version bump over GLM-5, but it fixes exactly the biggest hole: video timing rises from half a generation behind to 88, and GUI agents reach 95, among the highest single-dimension scores on this board. Composite 92.4 enters tier S — and the entire training stack is still Ascend. In tenders that require domestic chips, this is currently the only model satisfying chip, compliance and tier-S capability at once.
Last updated Oct 18, 2026Editorial scoring
Bottom line
Eight-dimension breakdown
editorial mapping · 0–100- GLM-5.392.4
Weights: perception 20 / documents 18 / charts 14 / video 14 / agents 14 / generation 6 / Chinese 6 / ecosystem 8. Definitions and sources on the methodology page.
Public benchmarks
as reported, not converted| Benchmark / metric | Value | Note |
|---|---|---|
| SuperCLUE-VLM 2026-10 | 89.4 | Official; up from GLM-5 |
| SWE-Bench Pro | 73% | GLM-5.3 official disclosure |
| 上下文 | 400K | Doubled from GLM-5’s 200K |
Public scores come from different versions and prompts and are not interchangeable with our composite; they are one evidence input.
Scenario mapping
five stress tests, one scale| Scenario | Score | Observation |
|---|---|---|
| Chinese receipt OCR | 94 · Stable on official forms and receipts | |
| Chart data extraction | 90 · Reliable on standard charts | |
| 30-min video QA | 88 · Gap closed; still behind Doubao | |
| Mobile GUI control | 95 · Among the best recovery rates here | |
| Chinese cultural context | 95 · Solid on official register and allusions |
Strengths and limitations
Strengths
- GUI agents 95, the most disciplined function-call formats, and better mid-task recovery than its peers.
- Full Ascend training plus on-prem images clears every hard gate in domestic-chip tenders.
- Chinese official documents, long academic prose and policy text remain top tier.
Limitations
- No dimension is outright #1, so pure capability contests go to Qwen3.8-Max and Doubao.
- Video at 88 is a big jump, but the gain is supported only by vendor figures so far.
- Closed weights; global ecosystem and multilingual coverage still clearly trail the Qwen family.
Price and real cost
List price
¥9/M tokens
≈ ¥9 / 1M output tokens. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥5.4 (assuming input at 25% of output price).
Use your volumeAbout GLM-5.3
Is GLM-5.3 open source?
Licence: Commercial + on-prem licence (Ascend images included). No public weights; access is via vendor API or an on-prem licence.
How much does GLM-5.3 cost?
≈ ¥9 / 1M output tokens. It ranks #4 with a composite of 92.4. Real bills depend on resolution, tiling, caching and volume discounts.
What is GLM-5.3 best for?
Domestic-chip and public-sector on-prem projects; Complex agent workflows and tool orchestration; Chinese official documents, policy and research reports。
Why does GLM-5.3 score 92.4?
The composite weights eight dimensions: perception 20%, documents 18%, charts 14%, video 14%, agents 14%, generation 6%, Chinese 6%, ecosystem 8%. Its strongest dimension is Ecosystem, cost & availability; its weakest is Generation & editing.
Personal ranking · Not an official score
Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.