Zellige AI releases Tessera-27B-B-CJK, a Qwen3.8-27B finetune using compressed CJK for token efficiency

Zellige AI has released Tessera-27B-B-CJK, an open-weight coding agent finetuned from Qwen3.8-27B under the Apache-2.0 license. The model is designed to reduce token consumption in agent loops by thinking in a compressed shorthand that mixes English, Chinese words, and symbols, while keeping final outputs and tool calls in plain English and code.

Token Efficiency and Performance

The model demonstrates significant reductions in thinking tokens compared to the untuned base model:

• Thinking Token Reduction: Uses 0.50–0.60x the thinking tokens of the base model on average (specifically 0.496x for typical turns, with a 90% CI of 0.385–0.624).
• Token-Weighted Efficiency: Achieves 0.644x token-weighted usage.
• Reliability: Maintains 100% valid tool calls and 0 repetition loops across 88 matched turns, matching the base model's reliability.
• Long Context: Successfully handled 6 out of 6 valid tool calls on prompts ranging from 101K to 131K tokens.

Training and Availability

The model was trained using four short LoRA steps on Tinker, leveraging the base model's own outputs. The pipeline included rewriting thoughts into compressed CJK for seed SFT, applying a token-level anti-loop fix, followed by agentic SFT and a rejection-sampling pass.

• Weights: Merged bf16 weights are available on Hugging Face, along with LoRA, GGUF, and NVFP4 versions.
• Limitations: The model is optimized for agent loops. It scores lower and exhibits more loops in competition math and hard algorithmic code tasks. Only the bf16 model has been evaluated; GGUF and NVFP4 files remain untested.