Space Coast Labs · Model demo
Meet Hemingway.
A small language model, fine-tuned to write in one distinctive literary voice. We built it to show what fine-tuning actually changes — behavior, not just facts. The live chat is coming online; this page explains what will be behind it.
What it is
The model started as Qwen2.5-3B-Instruct — a three-billion-parameter open-weights model, small enough to run cheaply and general enough to be worth shaping. We fine-tuned it on about 1,200 curated examples of writing in the style of Ernest Hemingway, drawn only from public-domain texts, then quantized the winning version — the epoch-3 checkpoint — so it serves on modest hardware.
How we know it worked
The honest test of a fine-tune is whether it beats the cheap alternative: the same base model with a carefully written prompt. We ran a blind evaluation on 100 held-out tasks — a judge model compared the fine-tuned output against the prompted baseline without knowing which was which, and the fine-tune won the clear majority of comparisons. The voice lives in the weights, not in a prompt somebody has to remember to paste.
Why this matters to your firm
Prompting steers a model; fine-tuning moves it. When output needs a consistent voice, a strict format, or a behavior that prompting cannot hold at scale, a small fine-tuned model is often the cheaper, more reliable tool. That is the trade this demo demonstrates — and the same method applies to your own documents and tone, scoped to your data and put through the same blind test before anyone calls it working.
“The morning was cold and clear. He watched the road and waited without speaking.”
Live chat is coming online.
The model will run on our own server, on modest hardware — when it lands, this page streams its real responses, at whatever speed they honestly arrive. Until then, nothing here fakes it. Two working paths meanwhile: