AI BriefingZelligeAI Releases Tessera-27B-B-CJK-RL: RL-Updated Qwen3.8-27B Agent Solves More Math with 0.67x Thinking Tokens
·2026.11.07 08:23
Key point
The RL update improves hard math and code scores to match the base Qwen3.8-27B while maintaining 0.67x thinking token usage on agent tasks.
Details
- ZelligeAI released Tessera-27B-B-CJK-RL, an RL-updated version of their CJK-thinking Qwen3.8-27B coding agent, available with open weights under Apache-2.0.
- On a main hard set of 118 problems (AIME 2022-2024, LiveCodeBench, MATH level 5), the model scored 100, compared to 89 for the previous Tessera-27B-B-CJK and 98 for the base Qwen3.8-27B.
- On 64 new agent states from 16 coding tasks, the model used 0.67x the base model's thinking tokens (90% CI 0.53-0.85) while producing 256 of 256 valid tool calls.
- Training involved 8 RL updates on Tinker using per-rollout weighting to prevent long failed rollouts from dominating updates; update 8 was selected for release.
- The model is available in BF16 and quantized formats (UL-Q3_K_XL through Q8_K_XL) in GGUF, with Q3_K_XL at 13.15 GB; users of the previous NVFP4 checkpoint need to redownload config.json to fix a vLLM loading issue.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Thinking Machines AI introduces Inkling 2 Small, planning Apache 2.0 open weights release next weekLongCat-2.5 Launches with Native Multimodality and 1.6T Parameters for Agentic Coding