Hemmingway-1 writes without the cloud: open weights fit on a 24 GB graphics card

On EQ-Bench 4, an emotional intelligence leaderboard, Hemmingway-1 scored 1330 points and took third place. For comparison: in a snapshot of the same leaderboard from 20.07.2026, leader Claude Fable 5 scored 1349,5, while GPT-5.5 stood at 1325,8.

The score is assigned by a machine. The personas and scores in EQ-Bench 4 are generated by the judge model itself, without human review, which is why BenchLM presents this test for reference and excludes it from its weighted ranking formula.

Previously, natural-sounding prose came at frontier prices. Now the weights are openly available under Apache-2.0, and 27B fits on a single GPU.

One-command launch. After installing llama.cpp, all you need is `llama-server -hf bartowski/Altworld_Hemmingway-1-GGUF:Q4_K_M`. The same GGUF files work in LM Studio, Ollama, and Jan AI. The recommended Q4_K_M quant weighs 17,44 GB and fits on a 24 GB GPU, the lightest IQ2_XXS takes 8,88 GB, and the original weights come to 54,7 GB. On a server, vLLM runs the model: `vllm serve Altworld/Hemmingway-1 --max-model-len 262144`.

Without your own GPU, there is the vendor's OpenAI-compatible API. Input costs $0,24 per million tokens, output costs $0,90, and the model ID is `hemmingway-27b`. The API advertises a 32 768-token context, although the model itself handles 262 144. The app has a free tier without a card on file; paid plans start at $9 per month.

The vendor labels the model English-first and reports its results on its own leaderboards: CommunicationBench is 50 points above Fable 5.1 and GPT-6 Astra, while StoryBench is 504 points above the base Qwen3.8-27B.

The model is English-first, so the first independent runs on open weights will show how it handles Russian texts.