Decision ·
Recommend thinking budget 512 for the deep tier
From the record
- Context
- DT512: thinking budget 512 = 1024 quality at 0.55× time → recommended config
- Why
- CHEAPER-PARITY with 1024 at 0.55× time
- What followed
- becomes the "recommended deep configuration"
- Reversible
- yes
- Status
- kept
- Alternatives considered
- 1024; 256
- 1024 (slower), 256 (−0.028, CHEAPER-LOSS), Gemma + thinking (−0.028)
- 256 (loses), 1024, 3072, Gemma 26B
Linked facts · 3
- Recommend thinking budget 512 for the deep tier — why: CHEAPER-PARITY with 1024 at 0.55× time — outcome: becomes the "recommended deep configuration"Recommend thinking budget 512 for the deep tier — why: CHEAPER-PARITY with 1024 at 0.55× time — outcome: becomes the "recommended deep configuration"Date2026-10-09Kinda record, not a measurement (set, status, judge, n, seeds: not recorded)Sourcenot in the public repositories yetOpen the data file
- Recommend deep tier = cut 6 + Qwen3.6-35B-A3B + thinking budget 512 (why: DT512 CHEAPER-PARITY: = 1024 quality at 0.55× time)Recommend deep tier = cut 6 + Qwen3.6-35B-A3B + thinking budget 512 (why: DT512 CHEAPER-PARITY: = 1024 quality at 0.55× time)Date2026-10-09Kinda record, not a measurement (set, status, judge, n, seeds: not recorded)Caveatoutcome: Became row H everywhere; selected best of ≥ 8 deep configs on V-para (winner's-curse ≈ 0.01–0.02)Sourcenot in the public repositories yetOpen the data file
- Recommended deep config = cut 6 + Qwen3.6-35B-A3B UD-Q2_K_XL + thinking budget 512 (why: 512 = 1024 in quality at 0.55× time)Recommended deep config = cut 6 + Qwen3.6-35B-A3B UD-Q2_K_XL + thinking budget 512 (why: 512 = 1024 in quality at 0.55× time)Date2026-10-09Kinda record, not a measurement (set, status, judge, n, seeds: not recorded)Caveatoutcome: not in claim build; needs a 12 GB device and an app-side thinking capSourcenot in the public repositories yetOpen the data file