Altworld model releases

Hemmingway-1 is a 27B open-weights writing model from Altworld, released 2026-09-21 under Apache 2.0. A fine-tune of Qwen3.8-27B, text-only, with a 262,144-token context and up to 32,768 output tokens, purpose-built for the everyday writing people actually do — emails, workplace messages, notes and other communication rather than coding or reasoning. On Altworld's own blind CommunicationBench it ranks first (beating GPT-6 Astra by ~50 points) and leads a human-likeness test by ~26 points; it places third on the independent third-party EQ-Bench 4 and performs level with Kimi K3 on StoryBench (vendor-run figures except EQ-Bench 4). Distributed on Hugging Face with hosted access via hemmingway.io. A derivative fine-tune on a Qwen backbone, added for parity with other tracked derivatives (Bonsai-2-27B, Ember-1, ZDTaichu5.0); flag for Jake if writing-focused fine-tunes are out of scope.