Zellige AI open-sources Tessera-27B-B-CJK-RL: RL-boosted CJK-thinking coder cuts thinking tokens by a third, matches base on hard math

ghgi_ · reddit · 2026-11-07

Zellige AI released Tessera-27B-B-CJK-RL, an Apache-2.0 open-weights coding agent fine-tune of Qwen3.8-27B that thinks in compressed CJK, adding 8 RL updates on top of the earlier SFT version to fix its math/code weakness.

Results (xhigh effort, same-day, same seeds):

AIME + HMMT 2026 (63 problems, 2 attempts each): 108/126 vs 97 for the previous version; 64K-window cutoffs dropped from 23 to 11;
Hard set (118 problems from AIME 2022-2024, LiveCodeBench, MATH level 5): 100 vs 89 previously, now roughly level with the base Qwen's 98;
Agent turns (64 held-out states): typical turn uses 0.67x the base model's thinking tokens (90% CI 0.53-0.85), with 256/256 valid tool calls.

Training: 8 RL updates on Tinker with per-rollout advantage weighting so long failed rollouts don't dominate; update 8 was selected and confirmed on held-out tasks.

Notes: most token savings come from training the model to think less; CJK packing is a smaller phrase-level bonus. GGUF repo ships BF16 plus UL-Q3KXL through Q8KXL quants (Q3 is 13.15 GB). NVFP4 checkpoints downloaded before Nov 5 need a config.json redownload for vLLM.

Original post →

Zellige AI open-sources Tessera-27B-B-CJK-RL: RL-boosted CJK-thinking coder cuts thinking tokens by a third, matches base on hard math

More from Models