ThinkingCap-Qwen3.8-27B

ThinkingCap-Qwen3.8-27B is a fine-tune of Alibaba’s Qwen3.8-27B, already on this board, from BottleCap AI. The developer’s blog is dated 22 September 2026, the weights appeared on Hugging Face on 23 September and MarkTechPost reported it on 23-24 September. Like the base model it takes text and images and returns text, with a served context of 262,144 tokens and 27.8 billion parameters. The change is in how much it reasons before answering: across twelve benchmarks BottleCap reports 37.2 per cent fewer thinking tokens on average, for an accuracy of 85.8 per cent against the base model’s 86.6 — less than one point lost for about a third less output. Those are the developer’s own measurements and are not stored as scores. For anyone running an open reasoning model on their own hardware, fewer thinking tokens mean faster answers and more requests per GPU, which is the practical case for it. BottleCap publishes quantised builds (GGUF, FP8, NVFP4 and MLX) alongside the full weights. The licence needs reading: PolyForm Small Business 1.0.0 permits use by small businesses only, with a separate personal-use grant, and larger organisations must ask BottleCap for a commercial licence. There is no hosted API from the developer.