ThinkingCap-Qwen3.6-27B
#13 in Reasoning ModelsBottlecap Ai · v3.6 · 27b · 2× · last seen Sep 18, 2026
40
Momentum
ThinkingCap-Qwen3.6-27B is a Reasoning model released by BottleCap AI, created through Fine-Tuning of the base model Qwen3.6-27B (Qwen Team, 2026). The training objective was to significantly reduce the length of thinking traces ("Thinking Tokens") without compromising the response quality, style, or safety mechanisms of the base model. The model is officially provided on Hugging Face under the Apache-2.0 license and is designed as a vision-capable (multimodal) dense 27B-parameter model with a long context window, available via API through various providers (including Featherless, Requesty/Sference).
Momentum trend
21.06.19.09.
Features
| Key Benchmark (%) | GSM8K accuracy rises from 93.3% (base) to 96.5% while using 74.1% fewer thinking tokens |
| Context Window (Tokens) | 262,144 tokens (per third-party providers; inherited from base Qwen3.6-27B) |
| License | Apache 2.0 |
| Multimodality | Vision-language model (text and image input); official GGUF release includes mmproj vision projector |
| Platform | Hugging Face (model repo bottlecapai/ThinkingCap-Qwen3.6-27B), deployable via vLLM, SGLang, llama.cpp; also available via API providers such as Featherless and Requesty (Sference) |
| Price per 1M Tokens | $0.40 input / $3.00 output per 1M tokens (example provider Requesty/Sference) |
| Release Date | July 2026 (initial v1 release) |