A Qwen 3.6 27B derivative claims to cut thinking tokens by more than 90%
AppealSame4367 · reddit · 2026-07-23
A Qwen 27B derivative claims 90% fewer thinking tokens
A Reddit user spotted ProCreations/grug-27b on Hugging Face and says the benchmark claims are notable:
- It is presented as a Qwen 3.6 27B derivative
- The repo claims it can use more than 90% fewer tokens for the thinking part
- If true, the poster says it would make a 27B model running at 3 TPS on an old laptop feel much faster in practice
The user has not tested it yet, so the post is still a claim-plus-curiosity rather than a verified evaluation.
More from Models
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- Aidan Gomez: model's CoT said no tool calls right before making them — aidangomez · 2026-10-27
- Nat Lambert: Cohere's model is narrow RL training on internal infra ill-suited for open weights — natolambert · 2026-10-27
- OpenAI teases DOTS is coming to Europe 'soon' — mark_k · 2026-10-27
- Clifford Pickover challenges AI to decipher a message he encoded in Norse runes — pickover · 2026-10-27