![\"ThinkingCap](\"https://bottlecapai.com/assets/thinkingcap-gemma-4-26b/press-thinkingcap-gemma-4-26b.jpg\")
This is the fifth model in the ThinkingCap series. We reduced how much Google's gemma-4-26B-A4B-it thinks, without compromising the quality of its answers.
\nWe have published four token-efficient models so far: ThinkingCap-Qwen3.6-27B, ThinkingCap-Qwen3.8-27B, ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B and ThinkingCap: Qwen3.8-Flash-Next.
\ngemma-4-26B-A4B is Google's sparse mixture of experts — 26B parameters with about 4B active per token. It is the third model family we have applied this work to, after Qwen and NVIDIA. On this one we put extra effort into keeping the accuracy intact: thinking comes down 38.9% on average at a cost of only 0.15pp of accuracy, which is smaller than the run-to-run spread on most of the suite.
\n*Video comparison of the original model and the ThinkingCap model, prompted with the same input.
\nBoth models run on the same hardware and use the same inference settings.*
We start with gemma-4-26B-A4B-it. The goal is to spend fewer tokens getting to the answer, and change nothing else: preserve the knowledge, reasoning ability, answer quality, instruction following and safety behaviour.
\n![\"Accuracy](\"https://bottlecapai.com/assets/thinkingcap-gemma-4-26b/accuracy.svg\")
Accuracy per task family, gemma-4-26B-A4B-it against the ThinkingCap model
\nAccuracy barely moves in any family, and in two it moves up. Instruction following and agentic\ngains 0.65pp and maths gains 0.39pp; code is flat at 0.02pp; knowledge and reasoning costs\n0.51pp and long-context and multimodal 0.92pp. Averaged over all twelve benchmarks the change\nis −0.15pp.
\n![\"Mean](\"https://bottlecapai.com/assets/thinkingcap-gemma-4-26b/mean-thinking-tokens.svg\")
Mean thinking tokens per task family, gemma-4-26B-A4B-it against the ThinkingCap model
\nThe thinking is where the change is, and every family gets shorter. Long-context and multimodal\nreduces the most, 54.4%, then knowledge and reasoning at 44.4%, maths at 36.1% and code at\n27.6%; instruction following and agentic reduces the least at 17.2%. Averaged over the twelve\nbenchmarks it is 38.9%.
\n| Benchmark | Accuracy (base) | Accuracy (ours) | Δ accuracy | Mean thinking tokens | Δ mean thinking | Trunc. % | Looping % |
|---|---|---|---|---|---|---|---|
| Math | |||||||
aime26 | 86.20% ±0.86 | 86.72% ±0.98 | +0.52pp | 14766 → 8963 | -39.3% | 5.2 → 0.5% | 5.5 → 0.9% |
hmmt_nov25 | 81.51% ±1.06 | 80.89% ±1.19 | -0.62pp | 16462 → 10542 | -36.0% | 9.0 → 1.2% | 8.6 → 1.2% |
hmmt_feb26 | 72.92% ±1.08 | 74.20% ±1.16 | +1.28pp | 22005 → 14717 | -33.1% | 18.6 → 4.6% | 17.0 → 4.5% |
| Code | |||||||
livecodebench | 76.98% ±0.56 | 76.96% ±0.67 | -0.02pp | 11560 → 8375 | -27.6% | 6.8 → 1.3% | 2.2 → 0.3% |
| Knowledge & reasoning | |||||||
gpqa_diamond | 79.92% ±0.57 | 78.96% ±0.67 | -0.96pp | 7699 → 4284 | -44.4% | 0.3 → 0.1% | 0.3 → 0.1% |
mmlu_pro | 84.55% ±0.29 | 84.64% ±0.16 | +0.09pp | 3907 → 1989 | -49.1% | 0.0 → 0.0% | 0.1 → 0.0% |
mmmlu | 84.09% ±0.26 | 83.43% ±0.27 | -0.66pp | 1212 → 731 | -39.7% | 0.0 → 0.0% | 0.0 → 0.0% |
| Long-context & multimodal | |||||||
aa_lcr | 64.62% ±2.49 | 60.62% ±2.89 | -4.00pp | 6519 → 2655 | -59.3% | 4.0 → 0.0% | 3.6 → 0.4% |
longbench_v2 | 59.43% ±1.14 | 60.88% ±1.29 | +1.45pp | 9486 → 2934 | -69.1% | 7.7 → 0.5% | 5.9 → 1.2% |
realworldqa | 70.83% ±0.56 | 70.61% ±0.56 | -0.21pp | 819 → 533 | -34.9% | 0.0 → 0.0% | 0.0 → 0.0% |
| Instruction following & agentic | |||||||
ifbench | 76.83% ±0.80 | 78.67% ±0.73 | +1.83pp | 6470 → 5585 | -13.7% | 0.4 → 0.4% | 0.5 → 0.5% |
tau2 | 64.25% ±0.97 | 63.73% ±1.26 | -0.53pp | 4374 → 3466 | -20.8% | 0.0 → 0.0% | 0.0 → 0.0% |
| Average (12 benchmarks) | 75.18% | 75.03% | -0.15pp | 8773 → 5398 | -38.9% | 4.35 → 0.72% | 3.64 → 0.76% |
Every row is the original BF16 against the ThinkingCap model, run on the same\nharness with the same settings on the same questions. Accuracy is the mean over seeds, ± the\n95% confidence interval from the spread between them. Every cell in the last row is an average\nacross the twelve benchmarks, each counting once — so Δ mean thinking there is the mean of\nthe twelve reductions, not the ratio of the two token figures beside it, which is 38.5%. The\nsame distinction applies to the family bars above, which plot each family's mean tokens. On\nτ²-bench the thinking tokens are per task, summed across the rollout's turns, rather than per\nresponse. Column definitions are in Evaluation methodology.
\nEvery one of the twelve gets shorter. The deepest cuts are LongBench v2 (−69.1%), AA-LCR\n(−59.3%) and MMLU-Pro (−49.1%); the shallowest IFBench (−13.7%) and τ²-bench (−20.8%). The two\nshallow rows are the two whose work is least about thinking at length: following a list of\nformat constraints, and applying a fixed customer-service policy through the right tool calls.
\nOn accuracy, most of the results are within noise: on eight of the twelve the gap is smaller\nthan the run-to-run spread, and five of the twelve come out ahead of the original model. The\nfour that move further than their intervals allow are IFBench (+1.83pp, in our favour), GPQA-Diamond\n(−0.96pp), MMMLU (−0.66pp) and AA-LCR (−4.00pp). AA-LCR is the one real cost in the suite.
\nLooping improves across the suite. The original model loops on 3.64% of its responses,\nagainst 0.76% for ours — and on the three benchmarks where\nit loops most, AIME 2026, HMMT Nov 25 and HMMT Feb 26, the rate falls from 5.5%, 8.6% and 17.0%\nto 0.9%, 1.2% and 4.5%.
\n![\"Accuracy](\"https://bottlecapai.com/assets/thinkingcap-gemma-4-26b/budget-curve.svg\")
Accuracy against the thinking-token budget, both models
\nRead left to right, this is the question a deployment actually asks: for a given token budget\nper question, how much accuracy do you get? ThinkingCap's curve sits above the original model's\nthrough the whole middle of the range — the same accuracy for a smaller budget — and the two\nconverge once the budget is wide enough that neither model is constrained by it.
\nReasoning models are noisy to evaluate. At the recommended temperature of 1.0 both the content\nand the length of a response move run to run, so a single seed tells you very little. Everything\nbelow is multiple seeds per benchmark, with accuracy reported as the mean and a 95% confidence\ninterval across seeds.
\nBoth sides are the same run. The original model and ours were evaluated by the same harness,\nwith identical sampling and serving settings. The original model was measured by us, not quoted\nfrom published figures. The configuration:
\nbase → ours.You can find the model here: bottlecapai/ThinkingCap-gemma-4-26B-A4B-it, with GGUF builds published alongside it.
","headings":[{"level":1,"text":"ThinkingCap: Gemma-4-26B-A4B — up to 69% and on average 39% less thinking, at unchanged accuracy","id":"thinkingcap-gemma-4-26b-a4b-up-to-69-and-on-average-39-less-thin"},{"level":2,"text":"TL;DR","id":"tl-dr"},{"level":2,"text":"Our fifth ThinkingCap","id":"our-fifth-thinkingcap"},{"level":2,"text":"The training objective","id":"the-training-objective"},{"level":2,"text":"Results","id":"results"},{"level":2,"text":"Evaluation methodology","id":"evaluation-methodology"},{"level":3,"text":"What the columns mean","id":"what-the-columns-mean"},{"level":2,"text":"Get the model","id":"get-the-model"}]}}