Qwen 3 Achieves Top Score in Human-Evaluated Summarization Benchmark for 30B Models
English summary
A Reddit user conducted a summarization benchmark using human-annotated summaries and an LLM judge. Among models in the 30B parameter range, Qwen 3 achieved the highest score, outperforming Gemma 4, which ranked second. The user speculated that newer Qwen versions might be increasingly optimized for agentic tasks, potentially impacting pure summarization performance, though Qwen 3 still led in this real-world annotation evaluation.
Chinese summary
一位 Reddit 用户利用人工标注的摘要和 LLM 裁判对模型进行基准测试。在约 30B 参数规模中，Qwen 3 得分最高，优于排名第二的 Gemma 4。该用户推测新版 Qwen 可能更偏向智能体任务优化，从而影响纯粹摘要能力，但在此次真实标注评测中 Qwen 3 仍居榜首。
Key points
Benchmark used human-annotated summaries and an LLM judge for evaluation.
基准测试采用人工标注摘要和 LLM 裁判进行评估。
Qwen 3 ranked highest among 30B models, followed by Gemma 4.
在 30B 模型中 Qwen 3 排名第一，Gemma 4 紧随其后。
Speculation that newer Qwen models may trade off summarization quality for agentic task optimization.
推测新版 Qwen 模型可能在智能体任务优化与摘要质量之间存在权衡。