Tessera-27B-B-CJK: Zellige AI's Qwen Fine-Tune Using Shorthand for Token Savings
Zellige AI released Tessera-27B-B-CJK, an open-source Qwen fine-tune optimized for agent loops by using a compressed CJK shorthand for internal reasoning. While internal thinking uses this token-saving shorthand, the model's final outputs and tool calls remain in standard English and code. Token optimization in multi-turn agent loops is a crucial engineering challenge that directly impacts operational costs and latency. By drastically reducing thinking token overhead without degrading tool call reliability, this approach offers a promising direction for more efficient autonomous AI systems. Evaluations show that Tessera uses roughly 0.50x to 0.64x the base model's thinking tokens while maintaining 100% valid tool calls and zero repetition loops on coding tasks. The model was trained via a four-step LoRA process involving seed SFT with compressed CJK, anti-loop fixes, and rejection sampling.
## BACKGROUND
Large language models in reasoning or agent modes often generate an internal chain-of-thought before producing an answer, which significantly increases API costs and execution time. CJK (Chinese-Japanese-Korean) characters carry a high semantic density per token, making compressed multilingual shorthands an effective way to reduce the total token count during model reasoning.