ZelligeAI/Tessera-27B-B-CJK

Tessera-27B-B-CJK: Efficient Agentic Coding with Compressed CJK Thinking

Tessera-27B-B-CJK is a 27 billion parameter coding-agent model developed by ZelligeAI, built upon the Qwen3.8-27B architecture. Its primary innovation lies in its ability to process internal thoughts using a compressed CJK (Chinese, Japanese, Korean) shorthand, which drastically reduces the "thinking tokens" required per agent turn. This optimization results in a 0.50-0.64x reduction in thinking tokens compared to the base model on held-out coding-agent tasks, while still producing answers and tool calls in plain English and code.

Key Capabilities

• Reduced Thinking Token Usage: Achieves 0.50-0.64x fewer thinking tokens per agent turn, leading to more efficient agentic workflows.
• Agentic Coding Specialization: Fine-tuned for coding-agent loops, maintaining 100% valid tool calls and zero repetition loops, similar to its base model.
• Long Context Handling: Supports long-context prompts up to 131K tokens, demonstrating robust performance with valid tool calls in extended scenarios.
• Compressed CJK Reasoning: Utilizes a unique compressed CJK format for internal reasoning, optimizing token efficiency without compromising output quality.

Good for

• Coding Agent Development: Ideal for applications requiring efficient, multi-step coding agents.
• Optimizing Token Usage: Suitable for scenarios where reducing computational cost and inference time for internal reasoning is critical.
• Complex Problem Solving: Designed for reasoning_effort="xhigh" settings in agentic loops, making it effective for intricate coding challenges.

This model is released as a beta, with its rank-32 LoRA merged into the base model's bf16 weights. While highly efficient for agentic coding, it's noted that for hard math and competitive programming, the base model may perform better.