New ThinkingCap Model Reduces Token Usage in AI Responses

• •ThinkingCap gemma-4-26B-A4B reduces reasoning tokens by 38.9% on average.
• •The model maintains an accuracy of 75.0%, slightly lower than the base model's 75.2%.
• •The PRO version offers up to 87.6% fewer thinking tokens for enhanced efficiency.

The ThinkingCap gemma-4-26B-A4B model, released by BottleCap AI, reduces reasoning tokens by an average of 38.9% while maintaining accuracy close to the base model. The accuracy of the new model is reported at 75.0%, slightly below the base model's 75.2%. This model is designed to enhance efficiency without compromising the quality of responses, achieving reductions in thinking tokens ranging from 14% to 69% across various benchmarks. The model is a mixture-of-experts type, utilizing a sparse architecture with 26 billion parameters. The PRO version claims to use up to 87.6% fewer thinking tokens, offering even faster responses and lower costs. The model is positioned as a drop-in replacement for existing systems, maintaining the same inference settings as its predecessor. The release is part of a series aimed at improving token efficiency in AI models.