2026-09-28 · America/Los_Angeles · 社区动态 · #17

Qwen3-VL 8B on a laptop vs Opus 5.5 / Sonnet 5 / GPT-5.6 on 137 messy documents: beat GPT-5.6 on tax forms, lost badly on Indian date formats[R]

I benchmarked Qwen3-VL 8B Instruct (Q4_K_M, Ollama, M5 24GB, ~30s/doc) against Claude Opus 5.5, Sonnet 5 and GPT-5.6 Terra on: - receipts: CORD (Indonesia) and SROIE (Malaysia), 30 each - 20 scanned 1980s-90s invoices, answer keys human-verified - 32 real IRS forms, 4 damage levels (generated this week, so not in anyone's training data) - 10 synthetic Indian bank statements, 15 CUAD contracts Results (documents fully right): - Opus 89%, - Sonnet 85%, - Qwen 8B 59%, - GPT-5.6 Terra 57%. Qwen-specific findings: - W-2s: 21/32 fully right vs GPT-5.6 Terra 7/32 - Indian bank statements: 2/10.…

热度 45.8 / 100；排名与评分保留该期记录。