open-text-detector-qwen36-35b-a3b-oct10-gguf by Alice

Model Card

By Alice, published under apache-2.0, revision 59453abf26ce.

GGUF files of the Open Text Detector token classifier, the model that scores every paper on opentextdetector.com. It detects AI-written and AI-edited text in English research papers, at the level of single tokens and sentences. These files are the LoRA adapters of woog/open-text-detector-qwen36-35b-a3b-oct10 (final-moe, Oct 10, 2026) merged into Qwen/Qwen3.6-35B-A3B (revision 995ad96), converted with llama.cpp (commit 3e390af18d) and quantized. This is not a chat model. Text generation is not what it was trained for. The GGUF gives the final hidden state of every token; the included token head (tokenhead.safetensors, one linear layer from 2,048 to 2 outputs) turns each hidden state into a…

Read Alice's full model card

Open Text Detector, Oct 10 model, GGUF

Website: opentextdetector.com

GGUF files of the Open Text Detector token classifier, the model that scores every paper on opentextdetector.com. It detects AI-written and AI-edited text in English research papers, at the level of single tokens and sentences.

These files are the LoRA adapters of woog/open-text-detector-qwen36-35b-a3b-oct10 (final-moe, Oct 10, 2026) merged into Qwen/Qwen3.6-35B-A3B (revision 995ad96), converted with llama.cpp (commit 3e390af18d) and quantized.

This is not a chat model. Text generation is not what it was trained for. The GGUF gives the final hidden state of every token; the included token head (token_head.safetensors, one linear layer from 2,048 to 2 outputs) turns each hidden state into a probability that the token is AI-written.

Files

File	Type	Size
`otd-final-moe-oct10-BF16.gguf`	BF16	69.4 GB
`otd-final-moe-oct10-Q8_0.gguf`	Q8_0	36.9 GB
`otd-final-moe-oct10-Q6_K.gguf`	Q6_K	28.5 GB
`otd-final-moe-oct10-Q5_K_M.gguf`	Q5_K_M	24.7 GB
`otd-final-moe-oct10-Q4_K_M.gguf`	Q4_K_M	21.2 GB
`otd-final-moe-oct10-Q4_K_M-imatrix.gguf`	Q4_K_M with importance matrix	21.2 GB
`token_head.safetensors`	Token head: `weight` (2 × 2048) and `bias` (2), float32	16 KB
`otd_score.cpp`	Scoring program built against llama.cpp: reads token windows, returns p(AI) per token
`score_gguf.py`	Scores a text file with `otd_score` and prints sentence scores
`results.json`	Every number in the tables below, with the evaluation metadata

How to run it

Build llama.cpp with CUDA, then build the scorer against it: g++ -O3 -std=c++17 otd_score.cpp -I llama.cpp/include -I llama.cpp/ggml/include -L llama.cpp/build/bin -lllama -lggml -lggml-base -o otd_score
Score a text: python score_gguf.py MODEL.gguf token_head.safetensors paper.txt (needs transformers for the Qwen3.6 tokenizer and safetensors).

The scorer must reproduce how the model was trained, or the scores mean nothing:

Tokenize with the Qwen3.6 tokenizer, without special tokens.
Cut the text into windows of 510 tokens every 256 tokens, plus a last window that ends at the final token.
Feed each window twice, back to back (1,020 tokens, positions 0 to 1,019), and read the hidden states of the second copy only. The model is causal, so the second copy is where every token has seen the whole window.
p(AI) of a token = softmax(W h + b)[1] with the token head. Where windows overlap, average a token's probabilities.
A sentence's score is the mean p(AI) of its non-blank tokens.
Flag a sentence only above a cutoff set on human writing (the site uses the cutoff that flags 1% or 0.1% of sentences in human papers from 2019–2022). A score of 0.5 is not a meaningful cutoff.

Accuracy of each file against the 16-bit model

The first row is the 16-bit model as published on the site (PyTorch, BF16 weights, LoRA not merged). The other rows are the GGUF files, scored through llama.cpp on the same evaluation sets with the same windows and cutoffs. These are the same measures as the "Runs compared" tables on the site's Evaluation page.

All numbers are percentages unless the column says AUROC. Recall is measured at the cutoff that flags 1% of each set's own human sentences.

Small-edit recall: AI sentences in the insert-one, rewrite-one, rewrite-two and split-one edits of the three held-out sets (heldout-writers, t21-heldout-v2, wide1-heldout; 7,243 sentences), pooled.
Δ vs 16-bit: difference in small-edit recall from the first row, in percentage points, with a 95% interval from a paired bootstrap over papers (2,000 replicates).
Spill-over FPR: share of human sentences next to a small edit that get flagged. Lower is better.
Paragraph edits: AI sentences of the paragraph and section cohorts of t21-heldout-v2 and wide1-heldout.

Model	File size	Small-edit recall	Δ vs 16-bit [95% CI]	Spill-over FPR	Rewrite one	Split one	Paragraph edits	Fable co-write
BF16 PyTorch (16-bit baseline)	–	74.6	—	1.09	64.2	62.1	94.7	26.3
BF16	69.4 GB	74.3	-0.3 [-0.4, +0.2]	1.08	63.7	61.4	94.4	26.3
Q8_0	36.9 GB	74.4	-0.2 [-0.5, +0.1]	1.08	63.8	61.6	94.2	26.0
Q6_K	28.5 GB	74.7	+0.0 [-0.4, +0.3]	1.10	64.2	61.6	94.1	25.7
Q5_K_M	24.7 GB	74.3	-0.4 [-0.7, +0.2]	1.11	64.0	61.6	94.6	25.7
Q4_K_M	21.2 GB	74.4	-0.2 [-0.7, +0.4]	1.18	63.6	62.7	95.3	26.5
Q4_K_M-imatrix	21.2 GB	74.4	-0.2 [-0.5, +0.5]	1.12	63.9	61.8	94.9	25.1

Standard held-out sets

Model	heldout-writers	t21-heldout-v2	wide1-heldout	Insert one	Rewrite one	Rewrite two	Split one
BF16 PyTorch (16-bit baseline)	65.0	80.4	69.3	92.8	64.2	74.3	62.1
BF16	64.7	80.3	68.5	92.8	63.7	74.0	61.4
Q8_0	65.1	80.2	69.0	92.6	63.8	74.3	61.6
Q6_K	64.8	80.6	69.2	92.7	64.2	74.6	61.6
Q5_K_M	64.7	80.2	68.4	92.0	64.0	74.1	61.6
Q4_K_M	64.8	79.9	69.8	91.7	63.6	74.4	62.7
Q4_K_M-imatrix	64.6	80.2	69.0	92.4	63.9	74.2	61.8

Other evaluation sets

Model	Humanized paragraphs	Co-writing	Co-writing AUROC	Light polish (document AUROC)	Other kinds of writing	Screening set small edits, dev	Screening set small edits, test
BF16 PyTorch (16-bit baseline)	97.0	26.3	0.729	0.814	98.5	45.5	60.0
BF16	97.1	26.3	0.728	0.815	98.5	46.0	58.7
Q8_0	97.0	26.0	0.728	0.815	98.5	45.5	60.0
Q6_K	97.0	25.7	0.728	0.817	98.5	45.0	58.0
Q5_K_M	97.2	25.7	0.730	0.819	98.6	43.4	57.3
Q4_K_M	96.8	26.5	0.725	0.814	97.2	45.0	56.7
Q4_K_M-imatrix	97.1	25.1	0.725	0.813	98.2	47.1	58.7

Running cost and agreement (one NVIDIA H200; tokens per second counts both copies of each window)

Model	Engine	Peak GPU memory	Tokens per second	Mean abs. difference from 16-bit sentence scores	Δ spill-over FPR [95% CI]	SHA-256
BF16	llama.cpp (LoRA merged)	65.5 GiB	3,032	0.0017	-0.02 [-0.03, +0.03]	`a916a5ee860e132d…`
Q8_0	llama.cpp (LoRA merged)	39.1 GiB	6,005	0.0020	-0.01 [-0.03, +0.03]	`60258a5a9e168aee…`
Q6_K	llama.cpp (LoRA merged)	31.4 GiB	5,420	0.0027	+0.01 [-0.03, +0.04]	`34df04e65237066b…`
Q5_K_M	llama.cpp (LoRA merged)	28.0 GiB	5,626	0.0047	+0.02 [-0.01, +0.08]	`8f1198ad97aeef71…`
Q4_K_M	llama.cpp (LoRA merged)	24.7 GiB	5,702	0.0073	+0.09 [+0.05, +0.15]	`1985f7c3df482bc1…`
Q4_K_M-imatrix	llama.cpp (LoRA merged)	24.7 GiB	5,654	0.0049	+0.03 [-0.01, +0.08]	`ee173bd01006beed…`

License

Apache-2.0, the license of the base model and of the LoRA adapters.

Identity and Version

Repository: woog/open-text-detector-qwen36-35b-a3b-oct10-gguf
Publisher: Alice
Task: Token classification
Modality: Text
Library: Not stated by the source
Parameters: Not stated by the source
Languages: en
Revision: 59453abf26ce12dbb27e6439d454afe184cab4c6
First published: 2026-10-15
Last updated: 2026-10-16

Files and Weights

13 files, 202.2 GB in total. The weights are 8 files totalling 202.2 GB in gguf, safetensors.

Weights8 files · 202.2 GB

Configuration2 files · 12.9 KB

Documentation1 file · 7.6 KB

Other1 file · 4.4 KB

Repository1 file · 2.0 KB

Every file

File	Type	Size	SHA-256
imatrix.gguf	Weights	384.3 MB	69d389ac69d1
otd-final-moe-oct10-BF16.gguf	Weights	69.4 GB	a916a5ee860e
otd-final-moe-oct10-Q4_K_M-imatrix.gguf	Weights	21.2 GB	ee173bd01006
otd-final-moe-oct10-Q4_K_M.gguf	Weights	21.2 GB	1985f7c3df48
otd-final-moe-oct10-Q5_K_M.gguf	Weights	24.7 GB	8f1198ad97ae
otd-final-moe-oct10-Q6_K.gguf	Weights	28.5 GB	34df04e65237
otd-final-moe-oct10-Q8_0.gguf	Weights	36.9 GB	60258a5a9e16
token_head.safetensors	Weights	16.5 KB	2a22466f47ee
results.json	Configuration	10.7 KB	—
score_gguf.py	Configuration	2.2 KB	—
README.md	Documentation	7.6 KB	—
otd_score.cpp	Other	4.4 KB	—
.gitattributes	Repository	2.0 KB	—

License and Download

License: apache-2.0
Access: Open weights, no gate
Download size: 202.2 GB

Download from Alice

Released by Alice through its official repository on Hugging Face. Read the license.

Built From

Derived from Qwen/Qwen3.6-35B-A3B
Quantized from Qwen/Qwen3.6-35B-A3B

Memory Requirements

Precision	Weights in memory
As published	202.2 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About open-text-detector-qwen36-35b-a3b-oct10-gguf

Can I use open-text-detector-qwen36-35b-a3b-oct10-gguf commercially?

Yes. open-text-detector-qwen36-35b-a3b-oct10-gguf is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Token classification

stanford-deidentifier-base

Stanford AIMI

Stanford de-identifier was trained on a variety of radiology and biomedical documents with the goal of automatising the de-identification process while reaching satisfactory accuracy for use in production. Manuscript in-proceedings. These model weights are the recommended ones among all available deidentifier weights. This work was supported in part by the Medical Imaging and Data Resource Center (MIDRC), which is funded by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) of the National Institutes of Health under contract 75N92020D00021 and through The Advanced Research Projects Agency for Health (ARPA-H)

Open weights mit 512 tokens transformers

View model

Model · Token classification

bert-portuguese-ner

Luís Filipe Cunha

This model is a fine-tuned version of neuralmind/bert-base-portuguese-cased It achieves the following results on the evaluation set: This model was fine-tunned on token classification task (NER) on Portuguese archival documents. The annotated labels are: Date, Profession, Person, Place, Organization All the training and evaluation data is available at: http://ner.epl.di.uminho.pt/ The following hyperparameters were used during training: - learningrate: 2e-05 - trainbatchsize: 16 - evalbatchsize: 16 - lrschedulertype: linear - numepochs: 4 - Transformers 4.10.0.dev0 - Pytorch 1.9.0+cu111 - Datasets 1.10.2 - Tokenizers 0.10.3

Open weights mit 512 tokens transformers

View model

Model · Token classification

privacy-filter-nemotron-GGUF

LocalAI-io

GGUF conversion of OpenMed/privacy-filter-nemotron, a fine-grained PII token-classification model — a fine-tune of openai/privacy-filter on the nvidia/Nemotron-PII dataset. It labels every token with a BIOES tag over 55 PII categories (221 classes) in a single forward pass, then decodes coherent spans with a constrained Viterbi procedure — so it can be served locally with no Python as the encoder/NER tier of a PII redactor. Where the base openai/privacy-filter covers 8 coarse categories, this fine-tune trades multilingual breadth for category depth: 55 fine-grained English categories (first/last name, government IDs, financial, healthcare, vehicle, digital, …). For the full model…

Open weights apache-2.0 gguf

View model

Model · Token classification

punctuate-all

KREDOR

This is based on Oliver Guhr's work. The difference is that it is a finetuned xlm-roberta-base instead of an xlm-roberta-large and on twelve languages instead of four. The languages are: English, German, French, Spanish, Bulgarian, Italian, Polish, Dutch, Czech, Portugese, Slovak, Slovenian. precision recall f1-score support accuracy 0.98 84425503 macro avg 0.83 0.74 0.77 84425503 weighted avg 0.98 0.98 0.98 84425503

Open weights mit 514 tokens transformers

View model

Model · Token classification

privacy-filter-multilingual-GGUF

LocalAI-io

GGUF conversion of OpenMed/privacy-filter-multilingual, a multilingual PII token-classification model (a fine-tune of openai/privacy-filter). It labels every token with a BIOES tag over 54 PII categories (217 classes) across 16 languages, so it can be served locally with no Python as the encoder/NER tier of a PII redactor. For the full model description, label space, evaluation, limitations, and citations, see the source model card — this card only covers the GGUF packaging and how to run it. This GGUF uses a custom architecture, openai-privacy-filter, that is not (yet) part of 1. privacy-filter.cpp (recommended) — a small standalone GGML engine for exactly this model family, on stock…

Open weights apache-2.0 gguf

View model

Model · Token classification

unbiased-toxic-roberta-onnx

Protect AI

This model is a conversion of unitary/unbiased-toxic-roberta to ONNX format using the Optimum library. Trained models & code to predict toxic comments on 3 Jigsaw challenges: Toxic comment classification, Unintended Bias in Toxic comments, Multilingual toxic comment classification. Built by Laura Hanu at Unitary. The huggingface models currently give different results to the detoxify library (see issue here). All challenges have a toxicity label. The toxicity labels represent the aggregate ratings of up to 10 annotators according the following schema: - Very Toxic (a very hateful, aggressive, or disrespectful comment that is very likely to make you leave a discussion or give up on sharing…

Open weights apache-2.0 514 tokens transformers

View model