reports/gpqa-thinkon-32k.md

GPQA Diamond at a 32k thinking budget: the three rows that capped at 16k

Rig: RTX 5090 32GB (capsule), llama.cpp build 10371 (5d16e81dd), llama-server, batch 1 · Window: 2026-10-03 to 2026-10-04, inside the box's cheap-electricity hours · Harness: llm-bench-rig, lib/evals/gpqa.py, idle-queue item 60-gpqa-thinkon-32k, artefacts in results/<slug>-thinkon-32k/ · Data: dataset/gpqa_thinkon_32k.csv · Sibling: MATH-500 at a 32k budget

The same check as the MATH-500 32k pass, on a second test. GPQA Diamond is 198 graduate-level multiple-choice questions in biology, physics and chemistry. At the think-on regime of max_tokens 16384, three rows hit the completion cap on more than 10% of the questions. This pass gives exactly those rows 32,768 tokens and changes nothing else that affects the answers.

TL;DR

1. Doubling the budget added 3.0 to 7.1 points. Qwable-27B 78.28 to 85.35, Nemotron 3.5 Lightning 30B-A3B 60.61 to 67.17, Qwen3.6-27B 76.77 to 79.80.
2. Every answer that finished inside 16k got the same verdict and the same completion length at 32k (124/124, 153/153, 167/167). Every added point is a question that had hit the 16k cap.
3. Nemotron 3.5 Lightning still caps on 19.7% at 32k, so its 67.17 is a floor, not a ceiling. The other two cap at 6.6% and 5.1%.

Setup

• Regime: thinking on, max_tokens 32768, ctx 40960, greedy, zero-shot, one pass per question. The 16k cells used max_tokens 16384, ctx 24576. Each 32k gpqa.json carries bench gpqa-thinkon-32k (separate higher-budget pass over rows capped >10% at 16k).
• Task: GPQA Diamond, 198 questions (Idavidrein/gpqa, diamond subset), letter answer.
• Rows: the same three GGUF files as the 16k think-on cells.

The table

Model	Quant	GGUF	16k	16k cap	32k	32k cap	32k tokens / correct
Qwable-27B	Q4_K_M	16.5 GB	78.28	22.7%	85.35	6.6%	11,737
Qwen3.6-27B	Q6_K	22.9 GB	76.77	15.7%	79.80	5.1%	11,108
Nemotron-3.5-Lightning 30B-A3B	Q4_K_M	24.5 GB	60.61	37.4%	67.17	19.7% ▲	19,701

▲ 10% or more of the questions still hit the 32k cap: a floor, not a ceiling. Per-row correct/total, capped_count, parse_failures, reasoning_fallback_count, tokens_per_correct and the median completion length at both budgets: dataset/gpqa_thinkon_32k.csv.

Question by question

Model	finished at 16k	same verdict and length at 32k	hit the 16k cap	of those, correct at 32k	still capped at 32k
Qwable-27B	153	153	45	30	13
Qwen3.6-27B	167	167	31	15	10
Nemotron-3.5-Lightning 30B-A3B	124	124	74	26	39

Reads

1. The 16k GPQA cells for these rows measured the budget. Finished answers did not change; what changed was how many answers finished.
2. The median question is unaffected. The median completion length is identical at both budgets on all three rows (5,814.5, 5,709 and 7,598 tokens), because the median question finishes well inside 16k.
3. The same pattern as MATH-500 and this year's contests. On all three tests, a think-on score below the budget is mostly a count of chains that did not finish (MATH-500 at 32k, AIME and HMMT at 32k).

Limits

• One greedy pass per question, one card. The Wilson 95% half-width on 198 questions is about 5 to 7 points at these scores, so the Qwen3.6-27B gain (3.0 points, 6 questions) sits inside it; the per-question table above shows where it comes from.
• GPQA Diamond only for these three rows. Rows that capped under 10% at 16k were not rerun.

Sources: results/<slug>-thinkon/gpqa.json (16k) and results/<slug>-thinkon-32k/gpqa.json (32k) on the rig; per-question comparison from the matching gpqa_progress.json files (correct, capped, completion_tokens per question), read 2026-10-07.