2026年10月23日 —— 9分钟读数
思维能力：Gemma-4-26B-A4B——最高达69%，平均减少39%，准确性不变
TL；数据
这是"思考卡普"系列的第五个模型。 我们减少了Google的gemma-4-26B-A4B-it认为的多少，同时不损害其答案的质量。
- 思维信号减少69.1%，平均为38.9%，精确度为0.15pp
- 关于Google模型的第一个思考卡普，以及迄今为止任何思考卡普的最小精度成本
- 循环率从原模型的3.6%下降到我们0.8%
- 相同的采样设置(考虑温度1.0，顶 p0.95)，相同的服务堆栈(vLLM)，滴入替换
思考PRO：想要更高的效率吗？这个模式的PRO版本使用的思维符最多减少87.6%，平均为62%. 更快的答案和较低的推论成本。 联系我们企业@bottlecapai.com.
我们的第五个思维能力
迄今为止，我们已公布了四个具有象征效率的模型：Thinking Cap-Qwen3.6-27B, Thinking Cap-Qwen3.8-27B, Thinking Cap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B和 Thinking Cap: Qwen3.8-Flash-Next.
gemma-4-26B-A4B是Google稀疏的专家混合体——26B参数，每个符號有约4B活性。 这是继Qwen和NVIDIA之后，我们把这项工作应用到第三个模范家庭。 关于这一点，我们付出了额外的努力来保持准确性：思维平均下降了38.9%，其精确度只有0.15pp，这比大部分套房的运行间隔要小。
最初的模型和思维卡普模型的视频比较，也是由同样的输入所激发的。
两个模型运行在相同的硬件上，并使用相同的推论设置。
培训目标
我们从gemma-4-26B-A4B-it开始。 目标是减少获得答案的代币开支，而改变一切：保存知识、推理能力、回答质量、遵循指示和安全行为。
结果
每个任务家族的精确度，gemma-4-26B-A4B-it 与 Thinking Cap 模型相对应
准确度几乎不会在任何家庭内移动，在两个家庭内，精确度会上升。 指令和代理收益0.65pp和数学收益0.39pp；代码为0.02pp；知识和推理费用为0.51pp和长文本及多式联运0.92pp。
gemma-4-26B-A4B-it 与 Thinking Cap 模型相对应
思想是变化所在，每个家庭都会变短。 长文本和多式联运减少最多，为54.4%，然后知识和推理减少44.4%，数学减少36.1%，代码减少27.6%；后续教学和代理教学减少至少17.2%。 12个基准测试的平均数为38.9%。
每行都是最初的BF16对抗"思维卡普"模型，运行在相同的吊带上，在相同的问题上设置相同。 准确度是种子的平均值，± 95% 的置信间隔与种子之间的传播。 最后一行的每个单元格都是12个基准测试的平均数，每个单元格都计算一次——所以，这意味着认为十二个减少的平均值是存在的，而不是旁边两个象征性数字的比例，即38.5%。 同样的区别也适用于上面的家庭栏杆，这些栏杆是每个家庭的平均标志。 在"2"下方，思考符号是每个任务，概括在展开的转折，而不是每个响应。 栏定义见评价方法。
十二部中的每一部都变短了最深层的削减是Long Bench v2(−69.1%)、AA-LCR(−59.3%)和MMLU-Pro(−49.1%)；最浅的IF Bench(−13.7%)和 2(−20.8%)。 两条浅行是两人的作品最不注重长短思考：遵循格式限制列表，通过正确的工具调用应用固定客户服务政策。
在准确性方面，大部分结果都是在噪声中：在十二位中的八位上，空隙比跑向散射小，十二位中的五位先于原型出场。 超出间隔的4个是IF Bench(+1.83pp，有利于我们)、GPQA-Diamond(−0.96pp)、MMMLU(−0.66pp)和AA-LCR(−4.00pp)。 AA-LCR是套房中真正的成本。
套房的循环会改善最初的模型循环了3.64%的回复，而我们的回复是0.76%，在它循环最多的三个基准测试，AIME 2026、HMMT Nove 25和HMMT Feb 26，比率从5.5%、8.6%和17.0%下降到0.9%、1.2%和4.5%。
精确度与思考预算，两种模式
从左到右读，这就是一个部署实际提出的问题：每个问题一个特定的象征性预算，你能得到多少准确性？思考 Cap的曲线坐落在原来的模型之上，穿过整个范围的中间——对于一个较小的预算来说，精确度是一样的——一旦预算足够宽，两个模型都不受它的限制，两者就会汇合。
评价方法
理性模型很吵，无法评估在1.0的推荐温度下反应移动的内容和长度运行，所以一个单一的种子告诉你很少。 低于每基准测试的每件东西都是多个种子，准确度报告为平均值，种子间置信度为95%.
双方同跑。 最初的模型和我们的模型都由同样的绳索评估具有相同的采样和服务环境最初的模型是由我们测量的，没有引用已公布的数字。 配置：
我们的内部评估套房在VLLM上运行两个模型
- 上下文——每基准测试每回65 536个信使，每转2 字箱每回65 536个信使。
- 取样——思考；温度1.0，最高0.95，最高 k 64. 我们的继承Google推荐的预设不变，因此比较中没有任何部分是解码变化。
- 精确度 – 平均跨种子的平均值为 95% t 间距，称仅在采样种子发生变化时，答案会移动多远。
- 种子——64种，关于AIME 2026年、HMMMT 2025年11月和HMMT 2月2026年；32种，关于GPQA-Diamond和LiveCode Bench;16种，关于Real WorldQA、IF Bench和\2 bench;8种，关于AA-LCR和Long Bench v2;4种，关于MMLU-Pro和MMMMLU，它们的问题数足够大，因此问题取样不是有限的噪音。
- 问题集——AIME 2026(30个问题)、HMMMT 2025(30)、HMMMT Feb 2026(33)、LiveCode Bench(175)、GPQA-Diamond(198)、MMLU-Pro(12 032)、MMLU(9 996)、AA-LCR(100)、Long Bench v2(370)、RealWorldQA(765)、IF Bench(300)和\bench(278项航空、零售和电信任务)。
- 分数——在提取的答案上精确匹配，但LiveCode Bench除外，其生成的代码运行在问题的测试上；IF Bench，其程序验证器得分；以及'2-bench和AA-LCR，这些被判断。 \\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\
纵队是什么意思？
- 准确度(基数)/准确度(我们)——如上；准确度是它们之间的差别，以百分比计。
- QQ 意思是思考——头条：思考符号的变化，在基准测试之间总和，并按其问数除以，所以一个长的问题重超过一个短的问题。
- 平庸的思维符号是推理结束前的符号，平均在基准测试的每一个问题和种子上，写我们的基础。
- 截断%——其思维从未结束，得分为零的回复比例。 循环% — 一个基于压缩的循环探测器所捕获的共享，它会标出一个其文本压缩远超普通流派所做反应的响应。
拿模型来
您可以在此找到模型：bottlecapai/ Thinking Cap-gemma-4-26B-A4B-it, GGUF 建置同时发布。
联系我们
有什么问题吗？或你想要承受的工作量？
发邮件到 Hey@bottlecapai.com
思考PRO：想要更高的效率吗？这个模式的PRO版本使用的思维符最多减少87.6%，平均为62%. 更快的答案和较低的推论成本。 联系我们企业@bottlecapai.com.
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Oct 23, 2026 · 9 min read
ThinkingCap: Gemma-4-26B-A4B — up to 69% and on average 39% less thinking, at unchanged accuracy
TL;DR
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.
- Up to 69.1% reduction in thinking tokens, on average 38.9% measured across twelve benchmarks, for 0.15pp of accuracy
- The first ThinkingCap on a Google model, and the smallest accuracy cost of any ThinkingCap so far
- Looping drops from 3.6% of the original model's responses to 0.8% of ours
- Same sampling settings (thinking on, temperature 1.0, top_p 0.95), same serving stack (vLLM), drop-in replacement
ThinkingCap PRO: want even better efficiency? The PRO version of this model uses up to 87.6% fewer thinking tokens, 62% on average. Faster answers and lower inference costs at scale. Contact us at enterprise@bottlecapai.com.
Our fifth ThinkingCap
We 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.
gemma-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.
Video comparison of the original model and the ThinkingCap model, prompted with the same input.
Both models run on the same hardware and use the same inference settings.
The training objective
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.
Results
Accuracy per task family, gemma-4-26B-A4B-it against the ThinkingCap model
Accuracy barely moves in any family, and in two it moves up. Instruction following and agentic gains 0.65pp and maths gains 0.39pp; code is flat at 0.02pp; knowledge and reasoning costs 0.51pp and long-context and multimodal 0.92pp. Averaged over all twelve benchmarks the change is −0.15pp.
Mean thinking tokens per task family, gemma-4-26B-A4B-it against the ThinkingCap model
The thinking is where the change is, and every family gets shorter. Long-context and multimodal reduces the most, 54.4%, then knowledge and reasoning at 44.4%, maths at 36.1% and code at 27.6%; instruction following and agentic reduces the least at 17.2%. Averaged over the twelve benchmarks it is 38.9%.
Every row is the original BF16 against the ThinkingCap model, run on the same harness with the same settings on the same questions. Accuracy is the mean over seeds, ± the 95% confidence interval from the spread between them. Every cell in the last row is an average across the twelve benchmarks, each counting once — so Δ mean thinking there is the mean of the twelve reductions, not the ratio of the two token figures beside it, which is 38.5%. The same distinction applies to the family bars above, which plot each family's mean tokens. On τ²-bench the thinking tokens are per task, summed across the rollout's turns, rather than per response. Column definitions are in Evaluation methodology.
Every one of the twelve gets shorter. The deepest cuts are LongBench v2 (−69.1%), AA-LCR (−59.3%) and MMLU-Pro (−49.1%); the shallowest IFBench (−13.7%) and τ²-bench (−20.8%). The two shallow rows are the two whose work is least about thinking at length: following a list of format constraints, and applying a fixed customer-service policy through the right tool calls.
On accuracy, most of the results are within noise: on eight of the twelve the gap is smaller than the run-to-run spread, and five of the twelve come out ahead of the original model. The four that move further than their intervals allow are IFBench (+1.83pp, in our favour), GPQA-Diamond (−0.96pp), MMMLU (−0.66pp) and AA-LCR (−4.00pp). AA-LCR is the one real cost in the suite.
Looping improves across the suite. The original model loops on 3.64% of its responses, against 0.76% for ours — and on the three benchmarks where it loops most, AIME 2026, HMMT Nov 25 and HMMT Feb 26, the rate falls from 5.5%, 8.6% and 17.0% to 0.9%, 1.2% and 4.5%.
Accuracy against the thinking-token budget, both models
Read left to right, this is the question a deployment actually asks: for a given token budget per question, how much accuracy do you get? ThinkingCap's curve sits above the original model's through the whole middle of the range — the same accuracy for a smaller budget — and the two converge once the budget is wide enough that neither model is constrained by it.
Evaluation methodology
Reasoning models are noisy to evaluate. At the recommended temperature of 1.0 both the content and the length of a response move run to run, so a single seed tells you very little. Everything below is multiple seeds per benchmark, with accuracy reported as the mean and a 95% confidence interval across seeds.
Both sides are the same run. The original model and ours were evaluated by the same harness, with identical sampling and serving settings. The original model was measured by us, not quoted from published figures. The configuration:
- Harness — our internal evaluation suite, running both models on vLLM.
- Context — generation capped at 65,536 tokens per response on every benchmark, and per turn on τ²-bench.
- Sampling — thinking on; temperature 1.0, top_p 0.95, top_k 64. Ours inherits Google's recommended preset unchanged, so no part of the comparison is a decoding change.
- Accuracy — the mean across seeds with a 95% t-interval, which says how far the answer moves when only the sampling seed changes.
- Seeds — 64 on AIME 2026, HMMT Nov 2025 and HMMT Feb 2026; 32 on GPQA-Diamond and LiveCodeBench; 16 on RealWorldQA, IFBench and τ²-bench; 8 on AA-LCR and LongBench v2; 4 on MMLU-Pro and MMMLU, whose question counts are large enough that question sampling is not the limiting noise.
- Question sets — AIME 2026 (30 problems), HMMT Nov 2025 (30), HMMT Feb 2026 (33), LiveCodeBench (175), GPQA-Diamond (198), MMLU-Pro (12,032), MMMLU (9,996), AA-LCR (100), LongBench v2 (370), RealWorldQA (765), IFBench (300) and τ²-bench (278 tasks across airline, retail and telecom).
- Scoring — exact match on the extracted answer, except LiveCodeBench, where the generated code is run against the problem's tests; IFBench, scored by its programmatic verifiers; and τ²-bench and AA-LCR, which are judged. τ²-bench also runs against a simulated user and is reported as the unweighted mean over its three domains.
What the columns mean
- Accuracy (base) / Accuracy (ours) — as above; Δ accuracy is the difference between them, in percentage points.
- Δ mean thinking — the headline: the change in thinking tokens, totalled across the benchmark and divided by its question count, so a long question weighs more than a short one.
- Mean thinking tokens are the tokens before the reasoning closes, averaged over every question and seed of the benchmark, written base → ours.
- Truncated % — the share of responses whose thinking never closed, which score zero. Looping % — the share caught by a compression-based loop detector, which flags a response whose text compresses far beyond what ordinary prose does.
Get the model
You can find the model here: bottlecapai/ThinkingCap-gemma-4-26B-A4B-it, with GGUF builds published alongside it.
Contact us
Have any questions, or a workload you want this run against?
Send us a mail at hey@bottlecapai.com
ThinkingCap PRO: want even better efficiency? The PRO version of this model uses up to 87.6% fewer thinking tokens, 62% on average. Faster answers and lower inference costs at scale. Contact us at enterprise@bottlecapai.com.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.