AI smell can now be scrubbed locally, but don't retire the detectors just yet

Abacus.AI has released the open weights for Smaug Humanizer on Hugging Face. The model is 12B, based on gemma-4-12B, licensed under Apache 2.0, and comes with usage instructions for vLLM, Transformers, and llama.cpp. What it does is straightforward: take a piece of AI-written text and output a version with the same facts, numbers, names, and formatting, but with a more human-sounding tone. The project page calls it a humanizer and states its goal as “destroy AI-slop” and bring creativity back to the internet. The slogan is grand; look at the implementation first.

The pain point isn't the phrase 'AI flavor' itself; it's the scenario. If someone sends a draft to the cloud to be scrubbed, privacy is exposed one more time. One user, Mohammad Vazin, noted that a humanizer only makes sense if the draft never leaves the local machine, and that running a small Gemma locally is the real feature. That judgment is more down-to-earth than the slogan. Smaug provides GGUF; Q8_0 can fit on a 16GB GPU, and Q4_K_M can run on an 8GB GPU or a laptop. Open weights and local inference lower the barrier.

The training method is worth a look. Abacus.AI starts from human text h from 2021 or earlier, which is known to be human-written. Then it has an LLM rewrite each passage in its own style to get AI text x. The model learns the reverse direction: from x back to h. Doing this directly has a pitfall: LLM rewrites often only change a few words, and the model would learn to lightly edit. The team used filtering to discard such pairs: check whether it is a real rewrite, whether the layout is consistent, and whether facts, numbers, names, and citations are preserved. In the end they kept 3,676 pairs and trained LoRA on gemma-4-12B, r=64, alpha=128, for 2 epochs. After that they ran GRPO for 200 steps with group size 8, and the reward moves the rewrite closer to the original human text h and away from the AI draft x, while five deterministic gates lock down numbers, layout, length, writing system, and spelling. The figure shows the two-stage process.

There is a detail in this reward design. It protects facts, but it cannot protect the strength of a commitment. One user, Tevfik Bulent Ongun, gave an example: 'We may deliver by Friday' and 'We will deliver by Friday' have the same date, but readers' expectations are very different. Modal verbs and attribution need separate checking. The model card also warns that the gates cover numbers and layout, but rewriting may still change details, so you should read it before publishing. This warning isn't just politeness; it's a boundary of use.

Evaluation used Pangram and an LLM judge. In the official examples, three types of input—a local election recount, a toast essay, and an explanation of lunar phases—were judged by Pangram as 100% AI, and after rewriting were judged as 100% human. The figure shows the input and output for the toast essay. The scores look great, but they shouldn't be taken as a pass for all text. The model card states that long texts are harder, support is mainly for English, and some rewrites will still be flagged by detectors. Detectors were already unstable. One user said AI detectors have long been like flipping a coin; now that a free humanizer is on Hugging Face, if schools still use detection scores to accuse people, they need to publicly explain the basis. That complaint feels realistic.

Another easily overlooked problem is that sounding human and sounding like you are not the same thing. One user, Jatin Garg, said matching human reference text may reward a generic human voice; preserving the author's own rhythm needs separate testing. Another user, Shin, has a more direct approach: give the model your own past posts plus a list of banned phrases and let it draft, which is more useful than a post-hoc humanizer. Post-hoc scrubbing can hide slop, but it cannot add real opinions. What readers ultimately remember is the viewpoint and judgment; tone is just the outer layer.

How to use it. The model card provides a fixed prompt that requires preserving facts, numbers, names, citations, structure, and language, and outputting only the rewritten text. Recommended sampling parameters are temperature 1.0, top_p 0.95, top_k 64, and stop using

Original: https://x.com/bindureddy/status/2114938847704571975

Model card: https://huggingface.co/abacusai/Smaug-Humanizer