gpt-6IMAGE

27B fine-tuned model claims 72% win rate over GPT-6 Astra on CommunicationBench

A writing model fine-tuned from Qwen3.8-27B claims it writes more like a person, winning 72% on hard prompts versus 9% for GPT-6 Astra, though the leaderboard is self-reported.

Md Ismail Šojal 🕷️

@0x0SojalSecArchived 2026.09.22Originally posted 2026.09.22

27B fine-tuned model claims 72% win rate over GPT-6 Astra on CommunicationBench (image 1 of 3)

27B fine-tuned model claims 72% win rate over GPT-6 Astra on CommunicationBench (image 2 of 3)

27B fine-tuned model claims 72% win rate over GPT-6 Astra on CommunicationBench (image 3 of 3)

Md Ismail Šojal 🕷️ @0x0SojalSecOriginally posted 2026.09.22

Open original on X

A 27B model just beats GPT-6 Astra & Fable 5.1 by fifty points on CommunicationBench. - this model just claimed it writes more like a person than models 10× its size. - 72% win rate on hard asks vs 9% for GPT-6 Astra. - Full disclosure they put up front: CommunicationBench, Human-Likeness, and StoryBench - based on a Qwen3.8-27B fine-tune for emails, texts, and the awkward note you keep rewriting. - http://huggingface.co/Altworld/Hemmingway-1

x.com/0x0sojalsec/status/2102092757762474409Archived 2026.09.22 10:18