ThinkingCap Qwen3.6 27B

ThinkingCap Qwen3.6 27B is a 27-billion-parameter reasoning model from BottleCap AI, a Prague-based lab co-founded by word2vec author Tomáš Mikolov. It is a deliberately minimal fine-tune of Alibaba's Qwen3.6-27B, trained on a curated multi-domain problem set with an objective that rewards reaching the right answer efficiently rather than rewarding correctness alone, so the model learns to stop reasoning once it has enough to answer. BottleCap reports a 50% average reduction in thinking tokens and over 90% in the best cases, with a 45.8% macro-average reduction across twelve out-of-domain benchmarks and accuracy holding or improving, including LiveCodeBench rising from 80.7% to 84.3% and GSM8K from 93.3% to 96.5% while using 74.1% fewer thinking tokens. Because reasoning traces are billed as output tokens, the saving lands directly on cost and latency for workloads that call a thinking model in a loop, such as agent steps, classification, and extraction. Style, knowledge, and safety behaviour track the Qwen3.6-27B checkpoint closely, and the weights are published on Hugging Face under Apache 2.0.