gohumanize-open-humanizer

Python client and CLI for the GoHumanize Open Humanizer, an open Qwen3-4B fine-tune (Apache-2.0) that rewrites AI-styled English text into more natural human prose. Trained on 2,629 pairs built from public-domain books and modern US government writing.

pip install gohumanize-open-humanizer
open-humanizer "It is worth noting that the committee reached a consensus."

from gohumanize_open_humanizer import Humanizer

# Your own copy of the model, which needs no key:
#   ollama pull hf.co/gohumanize/gohumanize-open-humanizer:Q4_K_M
h = Humanizer(base_url="http://localhost:11434/v1",
              model="hf.co/gohumanize/gohumanize-open-humanizer:Q4_K_M")
print(h.humanize("It is worth noting that the committee reached a consensus."))

# Humanizer() with no arguments points at the endpoint behind the browser demo,
# which is rate-limited and needs OPEN_HUMANIZER_API_KEY. To try the model without
# installing anything, use the demo at https://gohumanize.ai/open-model

The hosted endpoint sleeps when idle: the first request after a quiet period waits for a GPU cold start, one to two minutes, and later calls take a second or two. The default timeout is 300 seconds to allow for that.

Environment variables OPEN_HUMANIZER_URL, OPEN_HUMANIZER_MODEL and OPEN_HUMANIZER_API_KEY set the same options. No dependencies beyond the standard library.

The model is an educational release: it is separate from the production models of GoHumanize.ai and makes no claim about AI detectors.

Links

Several rewrites, best one returned

On modern prose the model sometimes plays safe and hands the text back almost unchanged (about one try in six for version 2 of the model). Each call therefore asks the endpoint for five rewrites (it generates them in parallel, so the wait is the same) and returns the one that moved furthest from the input while keeping a sensible length. Endpoints that cannot generate several at once, such as Ollama, are asked again only when the rewrite is barely a rewrite.

Humanizer(samples=1)   # one request, no choosing

open-humanizer --samples 1 "It is worth noting that ..."