Tessera, an open 27B tuned Qwen coding model that thinks in compressed CJK (huggingface.co)

Tessera-27B-B-CJK-RL, a newly released coding-agent model based on Qwen3.8-27B, introduces significant advancements in handling complex coding tasks. This model uniquely processes information using compressed CJK—which combines English and Chinese shorthand—allowing for more efficient thinking per agent turn compared to its predecessor. The model has undergone eight updates of reinforcement learning focusing on math and coding, resulting in enhanced problem-solving capabilities, particularly in competitive scenarios. This model is expected to have a substantial impact on the AI/ML community, especially in the realms of coding and mathematical reasoning. Its ability to handle extensive coding tasks within a constrained token window improves its effectiveness, achieving a notable increase in accuracy on new competition problems while maintaining efficiency. With merged bf16 weights amounting to 55.6 GB and the capacity to generate responses in plain English and code, Tessera-27B-B-CJK-RL exemplifies a significant leap in using reinforcement learning approaches to refine model performance in real-world applications.