Zellige AI has released Tessera-27B-B-CJK, a new finetune of the Qwen3.8-27B model. This model is designed for agent loops and utilizes a novel approach of thinking in compressed CJK (Chinese, Japanese, Korean) characters to significantly reduce token usage per turn. While the base Qwen model spent an average of 137 tokens per turn on coding tasks, Tessera-27B-B-CJK typically uses between 0.50x to 0.64x of that amount. The model maintains valid tool calls and avoids repetition loops, and has shown robustness on prompts exceeding 100K tokens. AI
IMPACT This model's approach to token compression could lead to more efficient AI agents and reduced operational costs.
RANK_REASON Release of a finetuned open-weights model with novel efficiency techniques. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
