Introducing efficiency focused “Thinking CAP” model series
Home<br>Home
Founders
Founders<br>About
Blog
Contact
About<br>Blog<br>Contact<br>hey(at)bottlecapai.com
Introducing efficiency focused “Thinking CAP” model series.
TL;DR
One of our core aims at BottleCap is efficiency in AI. To fulfil this mission, we fine-tuned Qwen3.6-27B to reduce unnecessary reasoning while preserving answer quality.
The result:
46% fewer reasoning tokens on average
Comparable benchmark performance
Fewer reasoning loops and failure cases
Lower latency and inference cost
Shorter, more to the point answers
Across twelve out-of-domain benchmarks, the model produced nearly identical accuracy while using roughly half as many thinking tokens.
We are releasing the model publicly on HuggingFace under a permissive Apache 2.0 licence. Anyone can freely download the model and instantly replace their local Qwen model to save money & time.
Contact us if you want further optimizations to save even more: enterprise@bottlecapai.com
Your browser does not support the video tag.
Why do reasoning models overthink?
Reasoning models have changed our expectations around language model performance. Given enough time and enough tokens, they can solve problems that were previously inaccessible to older generations of models.
The downside is that they often think far longer than necessary. Even relatively simple questions can trigger thousands of reasoning tokens:
revisiting already established assumptions,<br>repeatedly reformulating the same argument,<br>getting stuck in loops,<br>spending more time explaining than solving,<br>unnecessary filler words and verbose style for thinking that users do not even read.
This behaviour improves benchmark performance in some settings, but it also introduces costs: higher latency, higher inference spend, lower throughput, increased energy consumption, more compute, and more opportunities for failure.
Long reasoning traces have gradually become associated with intelligence, when in reality they often represent inefficiency. We wanted to cut through all of that, on the machine level. The question we asked was simple: how much of the modern model’s reasoning is actually necessary?
The objective
Our goal was intentionally conservative: do not try to make the model smarter or teach it new capabilities. We wanted to preserve knowledge, reasoning ability, answer quality, conversational style, instruction following, and safety behaviour.
The only thing we wanted to change was the amount of computation spent reaching an answer. In other words: keeping the same model, and making sure it overthinks less.
Training approach
Starting from the Qwen3.6-27B (Qwen Team, 2026) base model, we trained on a curated set of problems covering multiple domains and difficulty levels. The training objective rewarded efficient reasoning rather than simply rewarding correctness.
Importantly, the intervention was designed to remain as non-invasive as possible. The resulting model behaves very similarly to the original checkpoint — same style, same capabilities, same knowledge — but with substantially shorter reasoning traces. The model learns to stop once it has enough information to answer confidently.
Evaluation methodology
Evaluating reasoning models is more difficult than evaluating standard language models. At the recommended sampling temperature of 1.0, output quality and length can vary substantially between runs, so single-seed numbers give an incomplete picture. To reduce noise, we evaluated using full benchmark datasets, five independent random seeds, and statistical significance testing across all comparisons.
We evaluated both in-domain tasks (held-out portions of datasets related to training) and out-of-domain tasks (designed to test generalisation). The suite covers scientific reasoning and math, knowledge-based question answering, long-context tasks, system-prompt adherence, coding and agentic tasks, multi-turn conversational behaviour, and safety and model guardrails.
Results
Mean thinking tokens and accuracy per benchmark, base vs. Ours. Thinking-token bars are scaled per benchmark.<br>Out-of-domain token efficiency
Benchmarkacc (base)acc (Ours)Δ acctok (base)tok (Ours)matched Δ%looping (base→Ours)<br>gpqa_diamond85.5% ±1.483.8% ±1.9-1.6pp10,7773,351-67.8%0.4% → 0.4%<br>supergpqa64.0% ±0.264.0% ±0.1-0.1pp8,2463,384-58.4%0.2% → 0.1%<br>mmlu_pro85.9% ±0.285.4% ±0.2-0.5pp3,4551,290-53.7%0.1% → 0.1%<br>mmlu_redux93.9% ±0.193.9% ±0.1+0.0pp947406-44.8%0.0% → 0.0%<br>ceval90.6% ±0.790.3% ±0.6-0.3pp1,279663-47.1%0.0% → 0.0%<br>HMMT88.0% ±3.784.7% ±3.7-3.3pp39,27727,388-38.0%0.0% → 0.7%<br>livecodebench80.7% ±0.684.3% ±1.0+3.6pp15,74410,158-41.1%2.0% → 2.2%<br>longbench_v262.6% ±3.660.2% ±1.7-2.5pp1,7651,091-39.1%12.4% → 5.0%<br>realworldqa82.4% ±0.781.9%...