THE SOURCE IN CONTEXT
The full GLM-5.3 model card, with serving routes and evaluation notes. Keep it separate from Flash and from quantized community conversions.
Which artifact does this page identify?
This is the full GLM-5.3 card, separate from Flash and from community quantizations. Its local-serving section includes a distinct Ascend route. Use the exact checkpoint and the selected conversion as separate identities. The companion recipe’s 743B total and 39B active labels belong to that recipe rather than a measured memory receipt.
Source: zai-org · pinned document
What is the correct reasoning configuration?
The publisher’s controls match the documented low, high and max reasoning-effort levels, defaulting to max. clear_thinking remains a separate chat-template option with a false default. Retaining the default effort matters for benchmark reproduction; enabling clear_thinking for a chat workflow should be recorded as its own configuration choice.
Source: zai-org · pinned document
How much of the published score can be reproduced from a headline?
The card reports 66.9 on DeepSWE v1.1 and 88.2 on Terminal-Bench 2.1. Their footnotes use different sampling, context and timeout settings. Terminal-Bench 3.0 is another protocol: three rollouts per task, isolated task containers, up to 600 turns and a ten-hour timeout. A score from one task is not a prediction of workstation throughput or another task’s accuracy.
Source: zai-org · pinned document
What should be retained before comparing results?
Keep the exact task revision, harness, tool policy, reasoning effort, maximum output and treatment of failures. Some published tasks change installation checks or use a fallback score when a run fails. Those decisions are part of the measurement. Read the footnote for the specific benchmark before comparing this card with a report from another author.
Source: zai-org · pinned document
What does the source specify?
These declarations belong to the cited document. A planning minimum, model limit and measured runtime allocation are different quantities.
| Reasoning effort | low / high / max; default max |
|---|---|
| HLE-with-tools context / output cap | 300000 / 163840 tokens |
| DeepSWE context / timeout | 400K tokens / 6 hours |
| Terminal-Bench 3.0 repeats / turns / timeout | 3 rollouts / 600 agent turns / 10 hours |
Reported figures and their conditions
Selected facts reported by the original author. Each row has its own configuration and measurement boundary. Run AI Here has not reproduced these tests.
| Configuration / workload | Metric | Reported value | Evidence |
|---|---|---|---|
| DeepSWE v1.1 | DeepSWE | 66.9 reported scorereported value | Original row ↗ |
| Terminal-Bench 2.1 | Terminal Bench | 88.2 reported scorereported value | Original row ↗ |
DeepSWE v1.1 · DeepSWE · conditions
| model | zai-org/GLM-5.3 |
|---|---|
| model_revision | ba32f8c19c8738e2c57c71798c601d477a7bf844 |
| hardware | Not specified |
| executed_runtime_revision | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| measurement_scope | publisher capability evaluation, not local speed |
| benchmark | DeepSWE v1.1 |
| harness | mini-swe-agent |
| temperature | 0.95 |
| top_p | 1.0 |
| context | 400K |
| timeout_hours | 6 |
| reasoning_effort | max |
Terminal-Bench 2.1 · Terminal Bench · conditions
| model | zai-org/GLM-5.3 |
|---|---|
| model_revision | ba32f8c19c8738e2c57c71798c601d477a7bf844 |
| hardware | Not specified |
| executed_runtime_revision | Not specified |
| input_tokens | Not specified |
| output_tokens | Not specified |
| repetitions | Not specified |
| measurement_scope | publisher capability evaluation, not local speed |
| benchmark | Terminal-Bench 2.1 |
| harness | Claude Code 2.1.207 |
| temperature | 1.0 |
| top_p | 1.0 |
| max_new_tokens | 65536 |
| timeout_hours | 6 |
| reasoning_effort | max |
AI-assisted source reading and original editorial synthesis; no independent model execution.