Efficiency focused "Thinking CAP" model series BottleCap released the Thinking CAP model series, a fine-tuned version of Qwen3.6-27B that reduces reasoning tokens by 46% while maintaining benchmark performance. The model aims to cut inference costs and latency by eliminating unnecessary reasoning loops and verbose thinking. Released under Apache 2.0 on HuggingFace, it is designed for efficient AI deployment. 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 https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.6-27B 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 mailto:enterprise@bottlecapai.com 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, - repeatedly reformulating the same argument, - getting stuck in loops, - spending more time explaining than solving, - 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 https://huggingface.co/Qwen/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 Out-of-domain token efficiency | Benchmark | acc base | acc Ours | Δ acc | tok base | tok Ours | matched Δ% | looping base→Ours | |---|---|---|---|---|---|---|---| | gpqa diamond | 85.5% ±1.4 | 83.8% ±1.9 | -1.6pp | 10,777 | 3,351 | -67.8% | 0.4% → 0.4% | | supergpqa | 64.0% ±0.2 | 64.0% ±0.1 | -0.1pp | 8,246 | 3,384 | -58.4% | 0.2% → 0.1% | | mmlu pro | 85.9% ±0.2 | 85.4% ±0.2 | -0.5pp | 3,455 | 1,290 | -53.7% | 0.1% → 0.1% | | mmlu redux | 93.9% ±0.1 | 93.9% ±0.1 | +0.0pp | 947 | 406 | -44.8% | 0.0% → 0.0% | | ceval | 90.6% ±0.7 | 90.3% ±0.6 | -0.3pp | 1,279 | 663 | -47.1% | 0.0% → 0.0% | | HMMT | 88.0% ±3.7 | 84.7% ±3.7 | -3.3pp | 39,277 | 27,388 | -38.0% | 0.0% → 0.7% | | livecodebench | 80.7% ±0.6 | 84.3% ±1.0 | +3.6pp | 15,744 | 10,158 | -41.1% | 2.0% → 2.2% | | longbench v2 | 62.6% ±3.6 | 60.2% ±1.7 | -2.5pp | 1,765 | 1,091 | -39.1% | 12.4% → 5.0% | | realworldqa | 82.4% ±0.7 | 81.9% ±1.2 | -0.4pp | 2,959 | 913 | -48.5% | 0.4% → 0.1% | | AA-LCR | 76.2% ±3.0 | 74.2% ±2.2 | -2.0pp | 2,455 | 1,337 | -45.5% | 3.8% → 1.8% | | llm-system-prompts | 80.6% ±1.2 | 81.5% ±1.8 | +0.9pp | 1,737 | 976 | -40.0% | 11.6% → 8.2% | | Claw-Eval | 87.0% ±1.9 | 84.4% ±1.2 | -2.6pp | 919 | 689 | -25.2% | — | | macro mean | -0.7pp | -45.8% | Settings. Base Qwen/Qwen3.6-27B vs bottlecapai/Qwen3.6-27B-Efficient “Ours” . 5 seeds per condition; thinking on; mean ± 95% CI across seeds. Decoding: temperature=1.0, top p=0.95, top k=20, min p=0.0. Max generation tokens: 100,000 for the general suite gpqa diamond, mmlu pro, longbench v2, realworldqa and AA-LCR; 250,000 for HMMT Nov 2025 ; 32,768 for supergpqa and livecodebench; 16,384 for ceval and mmlu redux; 15,000 for llm-system-prompts; 49,152 for Claw-Eval. matched Δ% is the average paired per-question difference in thinking tokens negative = reduction . looping% is the share of responses stuck repeating the same reasoning chain. macro is the equal-weight mean across benchmarks. Across all benchmarks: a -45.8% average reduction in reasoning tokens with an approximately 0.7 percentage-point average accuracy difference. The headline result is straightforward — roughly half the reasoning, with effectively unchanged performance. In-domain evals Holdout test splits of datasets whose train splits are part of the finetuning mix — quality retention on in-distribution tasks in contrast to the out-of-domain benchmarks above . | Benchmark | acc base | acc Ours | Δ acc | tok base | tok Ours | matched Δ% | looping base→Ours | |---|---|---|---|---|---|---|---| | gsm8k | 93.3% ±1.5 | 96.5% ±0.3 | +3.2pp | 3,175 | 648 | -74.1% | 0.0% → 0.0% | | arc challenge | 97.0% ±0.3 | 97.6% ±0.4 | +0.6pp | 966 | 335 | -51.5% | 0.0% → 0.0% | | arc easy | 99.3% ±0.2 | 99.4% ±0.2 | +0.1pp | 566 | 260 | -44.5% | 0.0% → 0.0% | | commonsense qa | 86.7% ±0.7 | 88.2% ±0.9 | +1.5pp | 1,118 | 273 | -64.1% | 0.0% → 0.0% | | openbookqa | 96.0% ±0.5 | 96.7% ±0.6 | +0.7pp | 858 | 248 | -59.5% | 0.0% → 0.0% | | qasc | 91.7% ±0.7 | 92.2% ±0.5 | +0.5pp | 1,258 | 348 | -61.9% | 0.0% → 0.0% | | sciq | 97.0% ±0.2 | 97.5% ±0.2 | +0.5pp | 766 | 276 | -48.3% | 0.0% → 0.0% | | macro mean | +1.0pp | -57.7% | Settings. 5 seeds per condition; thinking on; mean ± 95% CI across seeds. Decoding: temperature=1.0, top p=0.95, top k=20, min p=0.0 bottlecapai/Qwen3.6-27B-Efficient uses the base’s sampling . Max generation tokens: 15,000 for GSM8K; 8,192 for the MCQ sets. Data: GSM8K is the full 1,319-row test split; the MCQ sets are capped at 1,000 rows OpenBookQA = 500 and QASC = 926 are smaller, so full . acc — exact-match on the final answer GSM8K / last-letter multiple-choice match MCQ ; matched Δ% and looping% are as defined for the token-efficiency table above. With focused training on holdout portions of the target datasets, we achieve a -57.7% reduction in reasoning tokens with a 1.0% increase in accuracy. Production matters more than benchmarks Unlimited token budgets are not how language models are actually deployed. Real systems operate under constraints: latency targets, context limits, cost ceilings, and throughput requirements. Under those conditions, efficiency matters. A model that reaches an answer in 2,000 tokens will often outperform one that reaches the same answer in 5,000. Reasoning compression therefore becomes more valuable as deployment scale increases. Our model spends its token budget solving the problem rather than narrating the solution. A very welcome bonus: more to the point answers Initially, we aimed to keep the answer style and quality completely unchanged, so that the users would never realise they are using a different model without examining the thinking trace. However, we accidentally applied our shortening also to the text of the answer after the