The card says reasoning_effort defaults to max and recommends keeping that default for benchmark reproduction. **SWE-Marathon**: We evaluate GLM-5.3 using Claude Code 2.1.207 with maximum effort level, `temperature = 1.0`, `top_p = 0.95`, `max_new_tokens = 128000`, and a 1M-token context window. For `strip-clone`, the original anti-cheat checks used overly broad import detection that could reject valid implementations. We removed the affected checks and performed llm-based inspection instead to avoid false positives. For `parameter-golf` and `trimul-cuda`, changes to the NVIDIA wheels caused the Docker image builds to fail, so we added `--extra-index-url https://pypi.org/simple` to restore successful builds.
条件
GLM-5.3; exact score definition and unit unspecified. 2026-11-18按完整模型卡修正单位、指标/题集或协议标注,数字不变。
GLM-5.3 · SWE-Marathon (v1.1) · GLM-5.3模型卡;v1.1 · Scores and configurations
GLM-5.3 record for SWE-Marathon (v1.1) · GLM-5.3模型卡;v1.1. 1 score record for this release, including 1 exact value (may include historical records). Includes 2 source locations.
The card says reasoning_effort defaults to max and recommends keeping that default for benchmark reproduction. **SWE-Marathon**: We evaluate GLM-5.3 using Claude Code 2.1.207 with maximum effort level, `temperature = 1.0`, `top_p = 0.95`, `max_new_tokens = 128000`, and a 1M-token context window. For `strip-clone`, the original anti-cheat checks used overly broad import detection that could reject valid implementations. We removed the affected checks and performed llm-based inspection instead to avoid false positives. For `parameter-golf` and `trimul-cuda`, changes to the NVIDIA wheels caused the Docker image builds to fail, so we added `--extra-index-url https://pypi.org/simple` to restore successful builds.
Conditions
GLM-5.3; exact score definition and unit unspecified. 2026-11-18按完整模型卡修正单位、指标/题集或协议标注,数字不变。