BottleCap AI has released ThinkingCap-Qwen3.8-27B, a fine-tune of Alibaba's Qwen3.8-27B aimed at one problem: large reasoning models think for too long. It is the second model in BottleCap's ThinkingCap series, which trains existing open-weight models to reach the same answer with less internal monologue.
Across a spread of benchmarks, BottleCap reports that ThinkingCap-Qwen3.8-27B uses around 37.2% fewer thinking tokens than the base Qwen3.8-27B model. The trade-off in accuracy is small: macro accuracy drops from roughly 86.6% on the base model to about 85.8% on the fine-tuned version, a fall of under one percentage point. For teams paying by the token, that is a meaningful saving for a marginal accuracy cost.
The fine-tune keeps the base model's reasoning-effort controls, so developers can still choose low, medium or x-high settings depending on how much thinking a task needs. BottleCap has not altered the underlying architecture, only how much the model deliberates before answering, which keeps it a drop-in replacement wherever Qwen3.8-27B already runs.
This continues a trend in open-weight fine-tuning: rather than training new foundation models, smaller labs such as BottleCap are specialising existing ones for cost and latency. Given Qwen's open licence, ThinkingCap-Qwen3.8-27B is freely available to download and self-host, which matters for anyone running reasoning models at volume rather than through a metered API.
Source: MarkTechPost
