<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga" -->

---
title: Gemma-4-26B-A4B with 39-69% less thinking at same accuracy
description: A new fine-tune of Google's gemma-4-26B-A4B-it cuts how many tokens the model spends 'thinking' without hurting accuracy. Across twelve benchmarks, thinking...
canonical: https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Gemma-4-26B-A4B with 39-69% less thinking at same accuracy | daily.dev
og:description: A new fine-tune of Google's gemma-4-26B-A4B-it cuts how many tokens the model spends 'thinking' without hurting accuracy. Across twelve benchmarks, thinking...
og:url: https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga
og:image: https://api.daily.dev/og/posts/z4hDInqgA.png
og:image:alt: Gemma-4-26B-A4B with 39-69% less thinking at same accuracy
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Gemma-4-26B-A4B with 39-69% less thinking at same accuracy

**[Hacker News](https://daily.dev/sources/hn)** · 8 min read · 1 upvotes · 0 comments

## Summary

A new fine-tune of Google's gemma-4-26B-A4B-it cuts how many tokens the model spends 'thinking' without hurting accuracy. Across twelve benchmarks, thinking tokens drop 38.9% on average and up to 69.1% on LongBench v2, at an average accuracy cost of just 0.15 percentage points. It's a drop-in replacement using the same sampling settings and vLLM serving stack. Looping behavior also falls, from 3.64% of responses to 0.76%. The one real accuracy cost is on AA-LCR, down 4.00pp. A PRO version cuts thinking tokens by up to 87.6%, 62% on average.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://bottlecapai.com/post/thinkingcap-gemma-4-26b>

## Questions this post answers

### How much can you reduce thinking tokens in Google's gemma-4-26B-A4B-it model without losing accuracy?

A fine-tuned version called ThinkingCap-gemma-4-26B-A4B-it reduces thinking tokens by an average of 38.9% across twelve benchmarks, with cuts up to 69.1% on LongBench v2, at an average accuracy cost of only 0.15 percentage points. It uses the same sampling settings (temperature 1.0, top_p 0.95) and serves as a drop-in replacement on vLLM.

_Teams tuning reasoning-model inference costs can track efficiency techniques like this on daily.dev._

### Does reducing reasoning length in LLMs increase looping or repetition in outputs?

No, in this case it decreased it. The original gemma-4-26B-A4B-it model loops on 3.64% of responses, while the token-efficient ThinkingCap version loops on only 0.76%. On the three benchmarks with the highest looping rates (AIME 2026, HMMT Nov 25, HMMT Feb 26), rates fell from 5.5%, 8.6% and 17.0% down to 0.9%, 1.2% and 4.5% respectively.

_Developers weighing reasoning-model tradeoffs can follow benchmark results like these on daily.dev._

### Which benchmark showed the biggest accuracy drop when shortening chain-of-thought reasoning in gemma-4-26B-A4B?

AA-LCR showed the largest accuracy cost, dropping 4.00 percentage points, from 64.62% to 60.62%, while thinking tokens on that benchmark fell 59.3%. This was the only benchmark where the drop exceeded the run-to-run noise interval in the model's disfavor by a meaningful margin, out of twelve benchmarks tested.

_Evaluating tradeoffs between speed and accuracy for reasoning models stays easier when you track results on daily.dev._

## Similar posts on daily.dev

- [Google Updates Gemma 4: The best Small LLM is back \!\!](https://daily.dev/posts/google-updates-gemma-4-the-best-small-llm-is-back--tvolamyq7) · Medium · 1 upvotes · 0 comments
- [Google’s Gemma 4 shines on local systems – both big and small](https://daily.dev/posts/google-s-gemma-4-shines-on-local-systems-both-big-and-small-qpert58dk) · InfoWorld · 0 upvotes · 0 comments
- [Google's Gemini 3.5 Flash costs 3x the model it replaced, and the era of cheap AI is ending](https://daily.dev/posts/google-s-gemini-3-5-flash-costs-3x-the-model-it-replaced-and-the-era-of-cheap-ai-is-ending-ochkeykzb) · XDA Developers · 2 upvotes · 0 comments

---

Tags: [#llm](https://daily.dev/tags/llm), [#vllm](https://daily.dev/tags/vllm), [#gemma](https://daily.dev/tags/gemma)

[View this post on daily.dev](https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Gemma-4-26B-A4B with 39-69% less thinking at same accuracy","url":"https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga"},"datePublished":"2026-10-23T17:54:52.573Z","dateModified":"2026-10-23T17:55:34.099Z","description":"A new fine-tune of Google's gemma-4-26B-A4B-it cuts how many tokens the model spends 'thinking' without hurting accuracy. Across twelve benchmarks, thinking...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/3c2a61726f23c82e2548c780553cd75a?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/3c2a61726f23c82e2548c780553cd75a?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,vllm,gemma","timeRequired":"PT8M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Gemma-4-26B-A4B with 39-69% less thinking at same accuracy"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/gemma-4-26b-a4b-with-39-69-less-thinking-at-same-accuracy-z4hdinqga#faq","mainEntity":[{"@type":"Question","name":"How much can you reduce thinking tokens in Google's gemma-4-26B-A4B-it model without losing accuracy?","acceptedAnswer":{"@type":"Answer","text":"A fine-tuned version called ThinkingCap-gemma-4-26B-A4B-it reduces thinking tokens by an average of 38.9% across twelve benchmarks, with cuts up to 69.1% on LongBench v2, at an average accuracy cost of only 0.15 percentage points. It uses the same sampling settings (temperature 1.0, top_p 0.95) and serves as a drop-in replacement on vLLM. Teams tuning reasoning-model inference costs can track efficiency techniques like this on daily.dev."}},{"@type":"Question","name":"Does reducing reasoning length in LLMs increase looping or repetition in outputs?","acceptedAnswer":{"@type":"Answer","text":"No, in this case it decreased it. The original gemma-4-26B-A4B-it model loops on 3.64% of responses, while the token-efficient ThinkingCap version loops on only 0.76%. On the three benchmarks with the highest looping rates (AIME 2026, HMMT Nov 25, HMMT Feb 26), rates fell from 5.5%, 8.6% and 17.0% down to 0.9%, 1.2% and 4.5% respectively. Developers weighing reasoning-model tradeoffs can follow benchmark results like these on daily.dev."}},{"@type":"Question","name":"Which benchmark showed the biggest accuracy drop when shortening chain-of-thought reasoning in gemma-4-26B-A4B?","acceptedAnswer":{"@type":"Answer","text":"AA-LCR showed the largest accuracy cost, dropping 4.00 percentage points, from 64.62% to 60.62%, while thinking tokens on that benchmark fell 59.3%. This was the only benchmark where the drop exceeded the run-to-run noise interval in the model's disfavor by a meaningful margin, out of twelve benchmarks tested. Evaluating tradeoffs between speed and accuracy for reasoning models stays easier when you track results on daily.dev."}}]}
```