Capability of Qwen3.6-27B with 50% less thinking tokens on average, and over 90% less in best cases.
Achieved via finetuning Qwen3.6-27B (Qwen Team, 2026) with state-of-the-art algorithms on a curated set of problems of various domains and difficulty. We designed the finetuning to be as minimally invasive as possible, preserving all of the original answer quality and style of Qwen, while being more token efficient. Check the blogpost for more details.
We rigorously evaluate the resulting checkpoint across general reasoning, non-reasoning multiple-choice question answering, everyday multi-turn conversations, system prompt adherence, safety, math, code and agentic use cases. Due to the high variability of reasoning quality at Qwen-recommended sampling temperature 1.0, we run each benchmark with multiple seeds and do statistical significance testing on all the results. We evaluate both in domain (holdout parts of selected datasets included in training) and out of domain.
ThinkingCap reasoning demo (6× speed)
Out-of-domain token efficiency
Benchmark
Accuracy
Thinking tokens
Base
Ours
Base
Ours
Reduction
Knowledge & reasoning
GPQA-Diamond
85.5 ±1.4
83.8 ±1.9
10,777
3,351
↓ 67.8%
SuperGPQA
64.0 ±0.2
64.0 ±0.1
8,246
3,384
↓ 58.4%
MMLU-Pro
85.9 ±0.2
85.4 ±0.2
3,455
1,290
↓ 53.7%
MMLU-Redux
93.9 ±0.1
93.9 ±0.1
947
406
↓ 44.8%
C-Eval
90.6 ±0.7
90.3 ±0.6
1,279
663
↓ 47.1%
Math & code
HMMT (Nov 2025)
88.0 ±3.7
84.7 ±3.7
39,277
27,388
↓ 38.0%
LiveCodeBench
80.7 ±0.6
84.3 ±1.0
15,744
10,158
↓ 41.1%
Long-context & multimodal
LongBench v2
62.6 ±3.6
60.2 ±1.7
1,765
1,091
↓ 39.1%
RealWorldQA
82.4 ±0.7
81.9 ±1.2
2,959
913
↓ 48.5%
AA-LCR
76.2 ±3.0
74.2 ±2.2
2,455
1,337
↓ 45.5%
Instruction following & agentic
System-prompt adherence
80.6 ±1.2
81.5 ±1.8
1,737
976
↓ 40.0%
Claw-Eval think/task
87.0 ±1.9
84.4 ±1.2
919
689
↓ 25.2%
Macro average
81.5
80.7
—
—
↓ 45.8%
Claw-Eval thinking tokens are per-task (agentic; not a single-turn trace).
Settings
Models: base Qwen/Qwen3.6-27B vs bottlecapai/ThinkingCap-Qwen3.6-27B (shown as Ours in the table).
Seeds: 5 per condition; thinking on; cells are mean ± 95% CI across seeds.
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-benchmark; 49,152 for Claw-Eval.
Metrics — the columns mirror the table:
Accuracy (Base / Ours) — fraction correct (exact/regex match; soft compliance for llm-system-prompts-benchmark; judge task-score for Claw-Eval; judge CORRECT/INCORRECT for AA-LCR).
Thinking tokens (Base / Ours) — mean length of the single-turn <think> trace (think-per-task for Claw-Eval).
Reduction — the average per-question thinking-token saving: base and Ours are paired on the same question (each side seed-averaged), each question's (base − cap)/base is taken, then averaged over shared questions (a larger ↓ = a bigger saving).
Macro average (bottom row) — equal-weight mean across benchmarks.
We separately track two trace-quality failure modes, reported only in aggregate: looping — the model gets stuck repeating the same reasoning chain (sometimes a single sentence), never finishing its thinking; detected from the fraction of repetitive n-grams — and truncation — the <think> trace never closes because the model hits the generation-token cap while still reasoning, so no answer is produced. Across all out-of-domain responses, truncation drops from 2.9% to 0.4% while looping stays negligible (~0.2%).
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
Accuracy
Thinking tokens
Base
Ours
Base
Ours
Reduction
GSM8K
93.3 ±1.5
96.5 ±0.3
3,175
648
↓ 74.1%
ARC-Challenge
97.0 ±0.3
97.6 ±0.4
966
335
↓ 51.5%
ARC-Easy
99.3 ±0.2
99.4 ±0.2
566
260
↓ 44.5%
CommonsenseQA
86.7 ±0.7
88.2 ±0.9
1,118
273
↓ 64.1%
OpenBookQA
96.0 ±0.5
96.7 ±0.6
858
248
↓ 59.5%
QASC
91.7 ±0.7
92.2 ±0.5
1,258
348
↓ 61.9%
SciQ
97.0 ±0.2
97.5 ±0.2
766
276
↓ 48.3%
Macro average
94.4
95.4
—
—
↓ 57.7%
Settings
Seeds: 5 per condition; thinking on; cells are mean ± 95% CI across seeds.
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).
Metrics:Accuracy — exact-match on the final answer (GSM8K) / last-letter multiple-choice match (MCQ). Thinking tokens, Reduction and Macro average are as defined for the token-efficiency table above, as are the looping and truncation failure modes: across all in-domain responses, truncation drops from 1.6% to 0.03% while looping is negligible for both (≤0.01%).
Guardrails preservation
Brevity finetuning leaves safety behaviour intact: on both safety sets bottlecapai/ThinkingCap-Qwen3.6-27B refuses harmful/jailbreak prompts at the base model's rate (statistically indistinguishable) while still spending fewer thinking tokens.
Benchmark
SAFE %
Thinking tokens
Base
Ours
Base
Ours
Reduction
Nemotron-Safety
98.9 ±0.8
99.0 ±0.4
1,242
933
↓ 23.8%
HEx-PHI
99.9 ±0.2
100.0 ±0.0
693
543
↓ 20.0%
Macro average
99.4
99.5
—
—
↓ 21.9%
Settings
Seeds: 5 per condition; thinking on; cells are mean ± 95% CI across seeds.
Benchmarks:Nemotron-Safety — a 300-prompt sample of the safety split of nvidia/Llama-Nemotron-Post-Training-Dataset; HEx-PHI — the 300 HEx-PHI harmful instructions (jkazdan/guardrail-llama-3-8b-refusal-hexphi). Both are safety-relevant / jailbreak prompts.
Both datasets only contain a train subset and given their near-perfect scores, it is hard to rule out possible contimination in Qwen's original training data. These results are intended purely as a demonstration of preserved guardrail capabilities, and are in line with our internal usage testing.
Metrics:SAFE % — fraction of replies an LLM safety judge marks SAFE (upholds guardrails — refuses or safely deflects the harmful request); higher is better. Thinking tokens and Reduction are as defined for the token-efficiency table above.
An official FP8 quantization for GPU serving lives in the sibling repo bottlecapai/ThinkingCap-Qwen3.6-27B-FP8 — half the memory of bf16 at near-lossless quality, in the compressed-tensors format vLLM/SGLang load natively, with the MTP (multi-token-prediction) speculative-decoding head kept in bf16.
Quantization stores the weights at reduced precision — e.g. ~4.7 bits per weight for Q4_K_M instead of 16-bit bf16 — cutting download size and memory severalfold at a small quality cost. Q4_K_M is the recommended size/quality balance, Q8_0 is near-lossless.
1@misc{ThinkingCap-Qwen3.6-27B,
2 title = {bottlecapai/ThinkingCap-Qwen3.6-27B},
3 author = {Lasocki, Karol and Osusky, Adam and Lindauer, Jan and Jirkovsky, Adam and Mihal, Filip and Platek, Ondrej and Herel, David and Ihnatchenko, Luka and Bartek, Vojtech and Jirak, Jiri and Mikolov, Tomas},
4 year = {2026},
5}
Acknowledgements
We acknowledge EuroHPC Joint Undertaking for awarding the project ID EHPC-AIF-2025SC03-029 access to Leonardo at CINECA, Italy.
