Hemmingway-1: an open 27B tuned to write like a person, not an AI

Apache-2.0, 262K context, built on Qwen3.8. It tops GPT-6 Astra on human-likeness, quants are already live, and today's M5 Ultra runs it easily.

Nowline SEP 22 8:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Trained to answer, not to hedge

    Hemmingway-1 is a 27B model fine-tuned from Qwen3.8-27B for everyday writing — messages, emails, casual notes — with a 262K-token context and Apache-2.0 weights you can ship in a product. It's tuned to return direct, human-sounding text instead of the usual preamble-plus-three-options wrapper.

  • It out-humans the frontier where it counts

    Altworld reports Hemmingway-1 landing 50 points over GPT-6 Astra on CommunicationBench and 26 ahead on human-likeness, and clearing 'hard asks' — money, admin, awkward messages — 72% of the time versus 9% for Astra; it placed third on EQ-Bench 4. Rival models buried the actual answer in commentary over 90% of the time, so this means far less 'As an AI…' cleanup on anything user-facing.

  • Quants and a demo are already live

    Within a day the community shipped GGUF and GPTQ-Int4 builds plus MTP variants, and there's a free no-signup web demo. You can trial the voice tonight and run a quantized copy on modest local hardware instead of a rented H100.

  • A desk that runs 150GB models — shipping today

    The M5 Ultra Mac Studio starts shipping September 22 with up to 256GB of unified memory and an 80-core GPU; reviewers ran a 156GB DeepSeek V4 Flash fully on-device to build a mini-game in minutes and called it a local-AI powerhouse. Frontier-size open weights now run on one desktop — and a 27B like Hemmingway-1 fits with room to spare.