R10 Stage 1 Humanizer (Qwen3.5-9B)

Probably the world's best humanizer model in terms of actually sounding like a human. Takes AI-generated text and rewrites it to sound human-written. It's a LoRA fine-tune of Qwen3.5-9B-Base (erickkwangg/qwen35-9b-r10-stage1).

Numbers: GPTZero averages 0.537 on the random40 OOD test, with a 44.2% strict 3-criteria pass rate. That's +18.6 points over the prior 8B best. decode_mild sampling is wired in.

Heads up: the main thing this model does is improve voice, so it may need a second pass for the facts (it's a voice fine-tune, not a general-purpose model). With heavy proper nouns and dates and things, it sometimes invents stuff while rendering the names: an institution gets renamed, a year goes missing, or a placeholder appears out of nowhere. The output reads convincingly human, so if you care about accuracy, go one more layer with a vanilla LLM and have it copy the facts over with the text you get out of here. Voice survives that pass intact.

First call after cold start takes 60-90 seconds while the 9B base loads in 4-bit. Subsequent calls are fast.