Tessera-27B-B-CJK-RL uses compressed thinking for efficient coding agents

ZelligeAI has released Tessera-27B-B-CJK-RL, a specialized open-weight coding agent model built on the Qwen3.8-27B architecture. Published on November 6, 2026, this beta release introduces a unique internal reasoning mechanism that compresses thought processes using a mix of English, Chinese characters, and symbols. The model aims to maintain high performance in competitive mathematics and coding tasks while significantly reducing the number of tokens generated per agent turn.

What happened

The new model is an evolution of the previously released Tessera-27B-B-CJK. It undergoes eight additional updates of reinforcement learning focused specifically on math and code problems. This training process results in a model that solves more novel competition-style math problems than its predecessor. Crucially, it retains most of the efficiency gains associated with the original Tessera line, which was designed to think less per turn than standard base models.

The release includes several repository options for developers. The primary offering is a set of merged bf16 weights totaling 55.6 GB. For those preferring parameter-efficient fine-tuning approaches, a 1.09 GB PEFT adapter is available. The team also provides GGUF formats with five Unsloth-layout replica quantizations and an NVFP4 checkpoint, though these latter formats have not been fully evaluated. All files are licensed under Apache-2.0, matching the base model.

How it works

The core innovation lies in the model's internal monologue, or "think" phase. Instead of generating verbose English reasoning steps, Tessera-27B-B-CJK-RL uses a compressed shorthand mixing Chinese characters, English words, and symbols. This method allows the model to convey complex logical states in far fewer tokens. For example, a typical agent turn might use only 45 tokens to describe test suite results and linting status, whereas the base model required 158 tokens for the same state. Despite this compression, final answers and tool calls remain in plain English and code, ensuring usability for developers.

Training involved starting from the Tessera-27B-B-CJK LoRA adapter and applying eight rounds of reinforcement learning on the Tinker platform. Each update sampled six rollouts for specific math and code problems, optimizing for correctness and cleanliness while penalizing excessive length relative to the base model. The advantage of each rollout was distributed across its thinking tokens, preventing long sequences from disproportionately influencing the gradient updates. The final LoRA pairs were merged into the base model's bf16 weights, with verification showing minimal deviation from the unmerged adapter performance.

Key details

• The model achieves 108 correct solutions out of 126 attempts on AIME and HMMT February 2026 problems, an improvement over the previous version's 97.
• Thinking per agent turn is approximately 0.67x that of the base model, maintaining significant efficiency gains.
• The release includes merged bf16 weights (55.6 GB), a LoRA adapter (1.09 GB), and untested GGUF and NVFP4 variants.
• Training used eight RL updates with a learning rate of 2e-5, scoring rollouts based on correctness and length efficiency.
• Recommended inference settings include reasoning_effort="xhigh", temperature 1.0, top-p 0.95, and top-k 20.
• The model is built on Qwen/Qwen3.8-27B and uses the stock tokenizer and chat template.

Why it matters

For engineers building autonomous coding agents, token efficiency directly impacts latency and cost. By compressing internal reasoning into a dense multilingual shorthand, Tessera-27B-B-CJK-RL reduces the computational load per decision step without sacrificing output quality. This allows agents to operate within tighter context windows, such as the 64K-token limit tested, where the new model failed significantly fewer times due to truncation compared to its predecessor. This efficiency gain is critical for scaling agent workflows that require extensive back-and-forth tool use and verification.

Furthermore, the improved performance on competitive math benchmarks suggests better general reasoning capabilities. While the model is specialized for coding, the ability to solve complex mathematical problems often correlates with stronger logical deduction in software engineering tasks. Developers can leverage this enhanced reasoning for debugging, algorithm design, and complex system architecture planning, provided they configure the inference parameters correctly to trigger the compressed thinking mode.

What you can do

• Deploy the merged bf16 weights using vLLM 0.31 or transformers 5.17 for local inference.
• Configure chat templates with reasoning_effort="xhigh" to activate the optimized thinking process.
• Use temperature 1.0, top-p 0.95, and top-k 20 to match the evaluation conditions.
• Drop past "think" tokens from history during multi-turn conversations to save context space.
• Test the LoRA adapter if you prefer parameter-efficient loading over merging weights.
• Avoid relying on the untested GGUF or NVFP4 checkpoints for production environments until further validation.