Altworld's Hemmingway-1 Skips the Fluff and Delivers Paste-Ready Messages
Altworld's 27B fine-tune of Qwen3.8 tops GPT-6 Astra on everyday message writing and reads as human 26 points more often.
Takeaways
- Altworld released Hemmingway-1, a 27B Qwen3.8 fine-tune for everyday message writing, Apache-2.0 licensed.
- Beats GPT-6 Astra by 50 points on CommunicationBench across 80 blind matchups.
- Wins 72% of hard asks versus GPT-6 Astra's 9%; 26 points clear on human-likeness.
- Placed 3rd on public EQ-Bench 4, ahead of GPT-5.5, Opus 4.7 and Opus 4.8.
- Returns just the message text instead of options and preambles, unlike most competitors.
- 262K context, runs on vLLM/SGLang/Transformers; English-only, not for medical, legal, financial use.
General chatbots often surround a requested message with introductions, alternatives, and tone notes. Altworld’s model card presents Hemmingway-1 as a simpler option: ask for a text, email, or reply, and receive the copy itself. The release applies a 27B open-weight model to short, tone-sensitive communication, a common workload that conventional reasoning and coding benchmarks rarely measure.
Built for the compose box
| Model size | 27 billion parameters |
|---|---|
| Base model | Qwen3.8-27B |
| Maximum context | 262,144 tokens |
| License | Apache-2.0 |
| Distribution | Hugging Face weights, quantizations, hosted playground, Mac app, and Android app |
Altworld released the weights alongside a hosted playground and a public code repository. The Apache-2.0 license permits commercial use, modification, and redistribution subject to its terms.
Eighty prompts, pairwise judging
Altworld created CommunicationBench from 80 real-world writing requests. For each matchup, a judge model received two shuffled answers to the same prompt without model labels. Altworld says it ran comparisons in both orders and used a judge model distinct from the systems being evaluated, reducing label and ordering bias.
On Altworld’s reported scoring scale, Hemmingway-1 ranks above Fable 5.1 and leads GPT-6 Astra by 50 points. The published chart also places Kimi K3, GLM-5.3, Grok 4.6, and DeepSeek V4 Pro behind it.
Altworld then rescored the same matchups by asking which response appeared to have been written by a person. Hemmingway-1 finished 26 points ahead of the next model on that measure. Within the category labeled “hard asks,” Altworld reports scores of 72% for Hemmingway-1 and 9% for GPT-6 Astra. It also leads the reported categories covering work, money and administration, and difficult interpersonal messages.