Cosmos Humanizer

The open-source AI humanizer built to beat the detectors that actually matter. Paste AI-written text in, get back writing that reads like a person wrote it, with your facts kept. Not a prompt wrapper and not a paraphraser with a new name: a model trained from the ground up on human rewrites, and it shows on Pangram.

Cosmos is an 8B diffusion language model (built on iLLaDA-8B), fine-tuned on our own dataset of about 160,000 AI → human rewrite pairs on an NVIDIA RTX PRO 6000, released here as GGUF for llama.cpp.

In our testing, Cosmos is the only open-source humanizer that actually gets past Pangram. We tried the other open humanizer models: none of them built their own dataset, and none of them held up on the detectors that matter.

Created by Mhmadhassnz · GitHub · X

Results on AI detectors

One pass, no manual edits. Pangram is one of the hardest detectors to beat, and it is the one we measured in full:

Tip: run it once or twice. Diffusion sampling is random, so two runs of the same text give two different rewrites. If a result doesn't pass, just run it again with another seed - our 17/21 run was a single pass with seed 0, and the same settings with another seed landed in the same range.

Why a diffusion model

Almost every humanizer out there is built on the same kind of model as the AI that wrote your text in the first place: an autoregressive transformer like GPT, Llama or Qwen. Those models write strictly left to right, one token at a time, and every word is picked by looking only at the words before it. That is exactly the habit detectors are trained to catch: the steady, predictable, always-the-likely-next-word rhythm of left-to-right generation. Rewriting AI text with another left-to-right model mostly swaps that fingerprint for another one of the same kind.

So we didn't. Cosmos is built on iLLaDA, a masked diffusion language model, and it writes in a completely different way:

• It writes the whole passage at once, not word by word. Generation starts from a canvas where every position is a blank mask. Over many steps the model fills in the masks it is most sure about first, anywhere on the canvas, and refines the rest around them, the way a person drafts: key points first, then the words that connect them.
• It is fully bidirectional. Every word is chosen while seeing what comes before and after it, not only what came before. Sentences get built around their content instead of being pushed out one predictable token at a time.
• No left-to-right fingerprint. There is no single "next token" chain for a detector to follow, so the word choice and rhythm come out much closer to how people actually write.
• Built for rewriting. Because the model sees the whole source and the whole draft at the same time, it can restructure sentences freely while holding on to the names, numbers and facts that matter.

Put a diffusion writer together with our own 160k-pair human dataset, and you get text that reads like a person wrote it, not like a second AI tried to hide the first one.

Before → after

Try it yourself: a full example

This is a complete, unedited input and output from the 17/21 run. The output scored 100% human on Pangram. Paste the input into your detector of choice to see it flagged as AI, then paste the output.

Input (AI-written)

Microfinance has long been promoted as a powerful tool for reducing poverty in developing countries. The concept gained international recognition through Muhammad Yunus, who founded the Grameen Bank in Bangladesh in 1983 and was awarded the Nobel Peace Prize in 2006. By providing small loans to people without access to traditional banking, particularly women, microfinance institutions aimed to help borrowers start businesses and improve their livelihoods.

The evidence on its impact, however, is more nuanced than early advocates suggested. In 2015, the American Economic Journal: Applied Economics published six randomized evaluations of microcredit programs in countries including Mexico, Morocco, and India. The studies found modest increases in business activity but no transformative effects on average household income, health, or education. Moreover, in some regions, such as Andhra Pradesh in India in 2010, aggressive lending practices contributed to over-indebtedness and a severe repayment crisis.

Despite these limitations, microfinance continues to offer meaningful benefits when designed thoughtfully. Access to savings accounts, insurance, and digital payment services can help poor households manage risk and smooth consumption. By combining financial services with training and consumer protection, institutions can better support vulnerable communities. Ultimately, microfinance is most effective as one component of a broader strategy for inclusive economic development.

Output (Cosmos Humanizer, one pass) - 100% human on Pangram

The concept was developed and globalized by Muhammad Yunus, who founded Grameen Bank in Bangladesh in 1983 and the Nobel Peace Prize in 2006, for providing small loans to people without who were unlikely to have access to traditional banking services. Small loans are used to start a new business, to buy assets, to improve their livelihood.credit loans can be used to buy assets like livestock equipment, the or to start small businesses.

2015 The American Economic Journal: Applied Economics conducted six randomized evaluations of micro credit in countries of Mexico, Morocco, India (Andhra Pradesh), and Tanzania in 2010 and the. in the studies of micro credit programs in Mexico and found increases in household income were reported small. The research studies also found that some areas, aggressive lending practices may lead to over- indebtedness.

Financial services access to poor households can help poor households to reduce risk, build assets and mitigate shocks. There are financial services such as credit, savings, insurance remittance, and other financial services that are available to households through formal financial services. Financial services can be provided through banks, institutions, lenders and other financial services providers. These financial services can help vulnerable communities in the provision of financial services, along with training and consumer protection. For example, financial services provided through mobile money and mobile banking can help improve access to financial services through digital technology.

Raw output on purpose, so you see exactly what the model writes: it keeps the names, years and figures, rewrites the structure, and leaves the small slips a person would fix in a quick read. Always give the result a proofread for facts before you use it (here, for example, "Tanzania" is not in the source).

Files

How to run (llama.cpp)

Cosmos is a diffusion model (see why), so it runs with llama.cpp's diffusion CLI instead of the usual chat tools (docs):

llama-diffusion-cli -m illada-humanizer-step3250-q8_0.gguf -ngl 99 -ub 1024 -c 2048 --diffusion-block-length 32 --diffusion-steps 512 --diffusion-algorithm 4 --temp 0.8 --seed 0 -p "Rewrite this text so it reads like a person wrote it: <your text>"

Settings of the 17/21 run

Tips

• Feed one paragraph per call; very long inputs (200+ words) can make it loop.
• If a paragraph comes out flagged, run it again with a different --seed: one or two runs is usually enough.
• Keep --temp around 0.8: lower reads stiffer (more AI-like), higher drifts from the source.
• --diffusion-steps should be at least the number of tokens you expect back (about 1.4 × the word count).

Good to know

• English only. Trained on English text.
• Facts: it is trained to keep names, numbers and years, but like any rewriter it can slip. Proofread before use.
• Speed: a 200-word paragraph takes about a minute on an RX 7800 XT at Q8_0; smaller quants are faster.
• Training data: our own dataset of about 160k AI → human pairs across essays, news, blogs, science and how-to writing. The dataset and training recipe are not released.
• Training hardware: trained on an NVIDIA RTX PRO 6000 workstation GPU.

License

Apache 2.0, the same as the base model GSAI-ML/iLLaDA-8B-Base. Please use it responsibly: it is a writing tool, so follow the rules of your school, workplace or publisher on AI-assisted writing.

Citation

@misc{cosmos_humanizer_2026,
  title  = {Cosmos Humanizer: an open-source diffusion humanizer},
  author = {Mhmad Hassn Alhmood},
  year   = {2026},
  url    = {https://huggingface.co/Mhmadhassnz/cosmos-humanizer-gguf}
}