Altworld Launches Hemmingway-1 to Write Clean Messages
Altworld has launched Hemmingway-1, a 27B open-weight model that writes ready-to-paste messages without conversational fluff, outperforming major proprietary models on communication benchmarks.

Altworld has released Hemmingway-1, a 27-billion-parameter fine-tune of Qwen3.8 designed specifically for everyday message writing. Distributed under the Apache-2.0 license, the model aims to solve a common frustration with general-purpose chatbots: the tendency to surround requested text with conversational preambles, alternative options, and tone notes. Instead, Hemmingway-1 outputs only the direct, paste-ready message text for emails, texts, or replies.
In evaluations, the English-only model demonstrated strong performance against much larger proprietary systems. On Altworld's new CommunicationBench, which uses 80 real-world writing requests judged blindly by an independent model, Hemmingway-1 beat GPT-6 Astra by 50 points. It won 72 percent of difficult prompts compared to GPT-6 Astra's 9 percent, while scoring 26 points higher on human-likeness. The model also outperformed Fable 5.1, Kimi K3, GLM-5.3, Grok 4.6, and DeepSeek V4 Pro on this benchmark. Furthermore, Hemmingway-1 secured third place on the public EQ-Bench 4, ranking ahead of GPT-5.5, Opus 4.7, and Opus 4.8.
For practitioners, Hemmingway-1 offers a highly optimized workflow for short, tone-sensitive communication tasks that traditional coding or reasoning benchmarks fail to capture. The model features a 262K context window and runs on popular inference frameworks including vLLM, SGLang, and Transformers. Altworld has made the weights, quantizations, and a public code repository available on Hugging Face, alongside a hosted playground, a Mac app, and an Android app. However, developers should note that the model is restricted from medical, legal, and financial use cases.
This is our own summary of reporting by AlphaSignal



