Free
Qwen vs. Qwen: two model generations head to head
Qwen3.6-27B against Qwen3.8-27B on the same 20 agent tasks - both dense, both 27 billion parameters, both on the same endpoint. 744 scored runs, 19 protocols read move by move. As a PDF, free of charge.
- Pass rate 88.4 vs. 94.2 percent - and why the new generation loses nine of 38 cells
- Seven case stories with original moves: the new generation wins four, the old one three
- The token bill: +32% median, +80% mean - at practically the same number of moves
- A chapter on the measurement error this volume corrects in itself: the wrong opponent
PDF, 65 pages, English - no account needed, no payment.