Hemmingway-1
- Type
- model
- Venue
- Hugging Face (Altworld model release)
- Year
- 2026
- Source
- huggingface
- Access
- public
- Language
- en
- Added
- 2026-09-29
- Verified
- 2026-09-29
Summary
Hemmingway-1 (Altworld) is a 27B open-weight model fine-tuned from Qwen3.8-27B to write everyday text 'as if written by a human' — returning just the message itself instead of options and explanations. On Altworld's own CommunicationBench of 80 real writing requests it reportedly beats frontier models at exactly this workload, with apps and a playground alongside the weights.
Keywords
writing · chat · open-weights · fine-tune
Topics
writing, chat, open-weights, fine-tune
Research notes
- Method: Fine-tune of Qwen3.8-27B (27B, hybrid Gated-DeltaNet linear attention + full attention, 262K context) specialized for human-like everyday writing: returns only the requested text, no options/preambles/explanations.
- Key findings: CommunicationBench (80 real-world writing requests, blind pairwise judging): ranked above Fable 5.1 and 50 points ahead of GPT-6 Astra; 26 points clear on 'written by a person?' judgments; Wins on money/admin, work, hard asks (72% vs GPT-6 Astra 9% on hard asks); loses on hostile storytelling and long story turns; Released alongside a hosted playground, Mac + Android apps, and web version
- Limitations: Self-reported benchmark (CommunicationBench) with no independent replication; evaluation is narrow (everyday writing only). License terms are inconsistently reported across sources (CC BY-NC 4.0 vs Apache-2.0) — verify before commercial use.
- Published 2026-09-20. Community quants/abliterations already exist. The name riffs on Hemingway; the pitch is 'paste-ready messages.'