Altworld presents Hemmingway-1, a 27-billion-parameter, open-weight English text-generation model built on Qwen3.8-27B and released under Apache-2.0. Its model card lists a 262,144-token context window and instructions for running it with Transformers, vLLM, SGLang, Docker Model Runner, or compatible local applications; it is not currently deployed by an inference provider. The model is designed for everyday messages, emails, workplace notes, difficult requests, and conversational writing, with an emphasis on returning the requested text without extended preambles or multiple options. Altworld says it evaluated the model on 80 real writing requests through blind, shuffled pairwise comparisons, with a separate model acting as judge. In those company-reported tests, Hemmingway-1 beat the listed competing systems, including Fable 5.1 and GPT-6 Astra, and was 50 points ahead of GPT-6 Astra; it also finished 26 points ahead of the next model in a human-likeness comparison asking which response appeared to be written by a person. The reported category results favor it on money and administration, work, difficult requests, and persuasion, while it performs worse on hostile storytelling and long story turns. On a separate measure of whether the answer is buried under commentary, options, or notes, Altworld says several comparison models did so in more than nine out of ten replies. Hemmingway-1 placed third on the public EQ-Bench 4, within 12 points of the top model, and ranked level with Kimi K3 on StoryBench while scoring 504 points above the model used as its starting point. Altworld discloses that CommunicationBench, Human-Likeness, and StoryBench are its own benchmarks; EQ-Bench 4 is English-first, and the company warns that the model can sound certain while being wrong and should not decide medical, legal, or financial matters.
