Community · Oct 28, 2026, 20:52 UTC

Qwen 3.8 27B outperforms Nanbeige 4.3 33B on jobs per hour in a user's 4x R9700 test

In a single user's firearms-business auditing workload on 4x R9700 GPUs, Qwen 3.8 27B processed about 45% more jobs per hour than Nanbeige 4.3 33B with thinking off, while Nanbeige was about 25% slower in single-stream tests. Accuracy with thinking on was 86.6% for Qwen versus 82.4% for Nanbeige.

• LabAlibaba
• ModelQwen 3.8 27B
• ModelNanbeige 4.3 33B
• Topicbenchmark
• Topicresearch