The Commons Lab's Hemmingway-1 Beats GPT-6 Astra at Everyday Writing
A 27B fine-tune of Qwen3.8 built for everyday messages beats much larger models at sounding like an actual human being.
- Hemmingway-1 is a 27B open-weights fine-tune of Qwen3.8-27B focused on everyday writing.
- Beats GPT-6 Astra, Kimi K3, GLM-5.3 and Grok 4.6 on the team's CommunicationBench.
- Wins human-likeness pairwise matchups by 26 points over the next model.
- 262k token context, BF16, runs on vLLM, SGLang, Transformers and Docker.
- Licensed CC BY-NC 4.0, free for non-commercial use, commercial by agreement.
- Weak on hostile storytelling and long multi-turn narrative; English-first.
The Commons Lab, a Swiss research group, has released Hemmingway-1, a 27-billion-parameter open-weights model tuned to draft texts, emails, workplace notes, and other short messages. The model is a fine-tune of Qwen3.8-27B, with weights available through its model card.
Hemmingway-1 targets a common failure mode in general-purpose assistants: wrapping a requested message in introductions, alternatives, bullet points, and explanations. Its training objective favors concise copy that a user can edit or send directly.
A narrow benchmark lead
The team’s CommunicationBench covers 80 everyday writing prompts and uses blind pairwise comparisons scored by a separate model. The model card reports that Hemmingway-1 ranked first against several larger general-purpose systems.
| Evaluation | Method | Reported result |
|---|---|---|
| CommunicationBench | Blind pairwise comparisons across 80 everyday writing requests | First place, ahead of Fable 5.1, GPT-6 Astra, Kimi K3, GLM-5.3, Grok 4.6, and DeepSeek V4 Pro; 50 points above GPT-6 Astra |
| Human-likeness | Pairwise judging of which response appeared human-written | 26 points above the runner-up |
| Hard asks | CommunicationBench category covering difficult requests | 72% for Hemmingway-1 and 9% for GPT-6 Astra |
| EQ-Bench 4 | Public emotional-intelligence benchmark using its published harness |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
