Open-Weight Models Under Creative Commons Attribution-NonCommercial-ShareAlike 4.0

Model · Text classification

finbert

Prosus AI

FinBERT is a pre-trained NLP model to analyze sentiment of financial text. It is built by further training the BERT language model in the finance domain, using a large financial corpus and thereby fine-tuning it for financial sentiment classification. Financial PhraseBank by Malo et al. (2014) is used for fine-tuning. For more details, please see the paper FinBERT: Financial Sentiment Analysis with Pre-trained Language Models and our related blog post on Medium. The model will give softmax outputs for three labels: positive, negative or neutral. To get the scores for all three labels, pass topk=None to the classifier. This model was trained using the Financial PhraseBank dataset: The…

This model is an Italian sequence-to-sequence model fine-tuned from the IT5-large for the task of inclusive language rewriting. It has been trained to analyze and rewrite sentences in Italian to make them more inclusive (if needed). For example, the sentence I professori devono essere preparati (The professors must be prepared) is rewritten as Il personale docente deve essere preparato (The teaching staff must be prepared). The model has been trained on a dataset containing a total of 4705 pairs of sentences, each pair containing an inclusive and a non-inclusive sentence. The dataset has been split as follows: We also leverage a small set of synthetic data (generated using a set of rules)…

Model · Audio classification

wav2vec2-large-robust-24-ft-age-gender

audEERING GmbH

The model expects a raw audio signal as input and outputs predictions for age in a range of approximately 0...1 (0...100 years) and gender expressing the probababilty for being child, female, or male. In addition, it also provides the pooled states of the last transformer layer. The model was created by fine-tuning Wav2Vec2-Large-Robust Timit and For this version of the model we trained all 24 transformer layers. An ONNX export of the model is available from Further details are given in the associated paper and tutorial.

Model

layoutlmv3-base

Microsoft

LayoutLMv3 is a pre-trained multimodal Transformer for Document AI with unified text and image masking. The simple unified architecture and training objectives make LayoutLMv3 a general-purpose pre-trained model. For example, LayoutLMv3 can be fine-tuned for both text-centric tasks, including form understanding, receipt understanding, and document visual question answering, and image-centric tasks such as document image classification and document layout analysis. Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei, ACM Multimedia 2022. If you find LayoutLM useful in your research, please cite the following paper: The content of this project itself is licensed under the…

Model · Feature extraction

splade-cocondenser-selfdistil

NAVER

SPLADE model for passage retrieval. For additional details, please visit: This is a SPLADE Sparse Encoder model. It maps sentences & paragraphs to a 30522-dimensional sparse vector space and can be used for semantic search and sparse retrieval. First install the Sentence Transformers library: Then you can load this model and run inference. If you use our checkpoint, please cite our work

Model · Audio classification

wav2vec2-large-robust-12-ft-emotion-msp-dim

audEERING GmbH

Please note that this model is for research purpose only. A commercial license for a model that has been trained on much more data can be acquired with audEERING. The model expects a raw audio signal as input, and outputs predictions for arousal, dominance and valence in a range of approximately 0...1. In addition, it provides the pooled states of the last transformer layer. The model was created by fine-tuning The model was pruned from 24 to 12 transformer layers before fine-tuning. An ONNX export of the model is available from doi:10.5281/zenodo.6221127. Further details are given in the associated paper and tutorial.

Model · Token classification

wikineural-multilingual-ner

Babelscape

This is the model card for the EMNLP 2021 paper WikiNEuRal: Combined Neural and Knowledge-based Silver Data Creation for Multilingual NER. We fine-tuned a multilingual language model (mBERT) for 3 epochs on our WikiNEuRal dataset for Named Entity Recognition (NER). The resulting multilingual NER model supports the 9 languages covered by WikiNEuRal (de, en, es, fr, it, nl, pl, pt, ru), and it was trained on all 9 languages jointly. If you use the model, please reference this work in your paper: The original repository for the paper can be found at https://github.com/Babelscape/wikineural. You can use this model with Transformers pipeline for NER. This model is trained on WikiNEuRal, a…

This model is an Italian classification model fine-tuned from the Italian BERT model for the classification of inclusive language in Italian. It has been trained to detect three classes: - inclusive: the sentence is inclusive (e.g. "Il personale docente e non docente") - notinclusive: the sentence is not inclusive (e.g. "I professori") - notpertinent: the sentence is not pertinent to the task (e.g. "La scuola è chiusa") The model has been trained on a dataset containing: - 8580 training sentences - 1073 validation sentences - 1072 test sentences The data collection has been manually annotated by experts in the field of inclusive language (dataset is not publicly available yet). The model…

Model · Audio classification

wav2vec2-large-robust-6-ft-age-gender

audEERING GmbH

The model expects a raw audio signal as input and outputs predictions for age in a range of approximately 0...1 (0...100 years) and gender expressing the probababilty for being child, female, or male. In addition, it also provides the pooled states of the last transformer layer. The model was created by fine-tuning Wav2Vec2-Large-Robust Timit and For this version of the model we only trained the first six transformer layers. An ONNX export of the model is available from Further details are given in the associated paper and tutorial.

Model

t5-base-grammar-correction

Vennify AI

This model generates a revised version of inputted text with the goal of containing fewer grammatical errors. It was trained with Happy Transformer using a dataset called JFLEG. Here's a full article on how to train a similar model. pip install happytransformer

Model · Time series forecasting

patchtst-fm-r1

IBM Research

This model card is for the non-commercial, research version of PatchTST-FM-r1. Please also check-out the Apache-2.0 licensed IBM Granite version. PatchTST was originally released prior to the interest in creating pre-trained, zero-shot time series foundation models that were capable of state-of-the-art performance on out of sample datasets. PatchTST-FM (patched time-series transformer-based foundation model) essentially has the architectural simplicity of PatchTST, but differs in some crucial ways. Coupled with a revised training strategy and a significantly larger training corpus, we are able to train a model that achieves state-of-the-art results on GiftEval. The architecture incorporates…

Model · Audio classification

discogs-maest-10s-pw-129e

Music Technology Group (Universitat Pompeu Fabra)

MAEST is a family of Transformer models based on PASST and focused on music analysis applications. The MAEST models are also available for inference in the Essentia library and for inference and training in the official repository. You can try the MAEST interactive demo on replicate. MAEST is a music audio representation model pre-trained on the task of music style classification. According to the evaluation reported in the original paper, it reports good performance in several downstream music analysis tasks. The MAEST models can make predictions for a taxonomy of 400 music styles derived from the public metadata of Discogs. The MAEST models have reported good performance in downstream…

Model · Summarization

ptt5-base-summ-xlsum

Recogna NLP

PTT5 Summ is a fine-tuned PTT5 model to perform Abstractive Summarization in Brazilian Portuguese texts. This model was fine-tuned on the datasets: RecognaSumm, WikiLingua, XL-Sum, TeMário.pdf) and CSTNews. For further information, please go to PTT5 Summ repository. author="Paiola, Pedro H. and de Rosa, Gustavo H. and Papa, Jo{\~a}o P.", editor="Xavier-Junior, Jo{\~a}o Carlos and Rios, Ricardo Ara{\'u}jo", title="Deep Learning-Based Abstractive Summarization for Brazilian Portuguese Texts", booktitle="BRACIS 2022: Intelligent Systems", year="2022", publisher="Springer International Publishing", address="Cham", pages="479--493", isbn="978-3-031-21689-3"} This model was fine-tuned using the…

IdeaLens-Qwen3.5-9B is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl -o scores.jsonl…

Model · Tabular classification

rt-plurel

Stanford Tabular and Relational (STAR) Project

RT-PluRel is a Relational Transformer checkpoint pair for in-context / few-shot entity prediction over multi-table relational databases (no per-task gradient training). Same architecture and file layout as stanford-star/rt-j — drop-in replacement. Architecture: ~85.6M blocks, dmodel 512, 8 heads, dff 2048 · text columns embedded with all-MiniLM-L12-v2 (dtext 384). Each folder contains model.safetensors (weights) and config.json (dims + text-embedding model). The paper/ subdirectory preserves the earlier RT-PluRel release:.pt checkpoints (12 blocks, dmodel 256, dff 1024) pretrained on synthetic relational databases generated by PluRel, plus the continued-pretraining and fine-tuned RelBench…

Model · Fill mask

logun-base

Heitorrosa

This model is a fine-tuned version of Itau-Unibanco/NorBERTo-base on an unknown dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 2 - evalbatchsize: 4 - gradientaccumulationsteps: 16 - totaltrainbatchsize: 32 - lrschedulertype: linear - lrschedulerwarmupsteps: 500 - numepochs: 1.5 - mixedprecisiontraining: Native AMP - PEFT 0.20.0 - Transformers 5.16.1 - Pytorch 2.11.0+cu128 - Datasets 4.0.0 - Tokenizers 0.23.1

Model · Speech recognition

koelsch-wav2vec2-ipa

Chem

IPA phoneme recognition for Kölsch (Ripuarian German, Cologne). Fine-tuned from facebook/wav2vec2-xls-r-300m with a CTC head over a 48-symbol IPA inventory. This is also the forced-alignment model. Its frame-level CTC posteriors drive torchaudio.functional.forcedalign to produce Praat TextGrids with word and phone tiers — see notebook 9 in the repository. Trained during the CIF Tandem Fellowship at IfL-Phonetik, University of Cologne, on Alles Kölsch (Bhatt & Lindlar 1998) — 4,670 utterances, 4.5 hours of spontaneous Cologne dialect speech from 105 speakers aged 10–88 across 49 neighbourhoods. Kölsch is Ripuarian German. It has no public speech dataset and no standardised spelling: 94.1 %…

IdeaLens-ModernBERT-L-NoParaphrase is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl -o…

Model · Image segmentation

MicroStructFormer

Nianqi Deng

MicroStructFormer is a high-capacity Mask2Former model built on a Swin-Large backbone, designed for the semantic segmentation of ultrastructures in Electron Microscopy (EM) images. This model has been trained to accurately segment the following critical ultrastructural elements: Note: While the model architecture theoretically supports axon and myelin segmentation, this specific release is optimized and calibrated for the classes listed below due to dataset annotations. You can use this model easily via the transformers library. This Hugging Face model repository is associated with the following bioRxiv preprint: These model weights and associated architectures are provided for…

IdeaLens-Qwen3.5-9B-PerItem is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl -o…

IdeaLens-ModernBERT-L is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl -o scores.jsonl…

IdeaLens-ModernBERT-L-PerItem is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl -o…

IdeaLens-ModernBERT-L-RolesOnly is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl -o…

ProseLens-ModernBERT-L is a prose-level detector: it reads the document itself and judges who wrote the words. It is the ModernBERT counterpart of ProseLens and the comparison point for the idea-level detectors. Score documents directly; no outline is needed. The idealens package (PyPI) applies this model and the thresholds in this repo: Input is JSONL with a text field per document. thresholds.json holds this model's cuts at 0.1%, 0.5%, 1%, 2% and 5% false-positive rates, fitted on the 80,000 human documents of WildOutlines' calibration split: one global cut per rate, plus per-format cuts. The package applies them. A cut fitted for one model does not transfer to another model's scores. For…

Model

alephllm-mini-beatrix-training

AbstractPhila

Live training ground for AlephLLM: signed-address (aleph) language models. Code, presets, trainer, and the full test array live in the source repo — github.com/AbstractEyes/alephllm (pip install git+https://github.com/AbstractEyes/alephllm, package geolip.alephllm). This repo holds what training produces, one prefix - mini-beatrix-3 — COMPLETE (2026-10-05; started 2026-09-20, relaunched 2026-09-27 on the engineered form and 2026-09-28 on amoe-lora 0.2.11). The v3 craft: 32 blocks, d 1024, 16 heads, context 4096, byte vocabulary with the trigram byte embedding, about 376M parameters. Every block carries a CausalSplatHUB (four address constellations, K 64 at D 128, the exact chunked scan at…

Model

alephllm-nano-beatrix-training

AbstractPhila

The training ground for the nano line of Beatrix: byte-level crafts under ten million parameters, trained from birth to find out what a craft that small can hold, how it holds it, and how fast it trains. The crafts, the trainer and the model code live in the library github.com/AbstractEyes/alephllm (package geolip.alephllm); the bed that drives them lives in github.com/AbstractEyes/geolip-nano-beatrix-trainer. This repository holds what training produces. It is the small sibling of alephllm-mini-beatrix-training, where the 237M and 257M teachers were trained. Every run trains one small craft: six blocks of width 256, eight heads, a bank of seven experts of width 256 in every block, 7.7 to…

Model · Video text to text

EventLens-4B

Guowen Zhang

EventLens-4B is a video-language model developed through event-structure reinforcement learning and multi-teacher on-policy distillation (MT-OPD). Built on Qwen3-VL-4B-Instruct, it learns to recognize event boundaries, progression, and relations, alongside temporal grounding and video question answering. The model is part of EventLens: Event-Structure Reinforcement Learning for Video Understanding, developed by AIM3 Lab, Renmin University of China, and Kling AI. Selected results reported in the EventLens paper, compared with the corresponding pretrained backbone under the shared 256-frame protocol. QVHighlights, ActivityNet, and Charades-STA use mIoU. E.T. Bench and VinoGround use their…

Model · Video text to text

EventLens-8B

Guowen Zhang

EventLens-8B is a video-language model developed through event-structure reinforcement learning and multi-teacher on-policy distillation (MT-OPD). Built on Qwen3-VL-8B-Instruct, it learns to recognize event boundaries, progression, and relations, alongside temporal grounding and video question answering. The model is part of EventLens: Event-Structure Reinforcement Learning for Video Understanding, developed by AIM3 Lab, Renmin University of China, and Kling AI. Selected results reported in the EventLens paper, compared with the corresponding pretrained backbone under the shared 256-frame protocol. QVHighlights, ActivityNet, and Charades-STA use mIoU. E.T. Bench and VinoGround use their…

static quants of https://huggingface.co/ProsusAI/finbert For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here…

Model

grammar-de

Kotoshu

Fully local German grammar correction model for the ≤250 MB weight class (run anywhere, CPU real-time, no network). iterative apply-and-retag decoding, on a microsoft/mdeberta-v3-base encoder. byte-fallback), pruned by weight-truth-verified remap — int8 ONNX gate, 48.5% yield) from typed corruption packs (articles, verb agreement, V2, separable verbs, case government); two-stage differential-LR schedule, byte-identical with HF) + ONNX Runtime; inference is milliseconds per sentence on a laptop CPU. German has no public GEC benchmark we can license for automated scoring, so we report a held-out judge study instead: This is a deliberately conservative corrector: it almost never invents…

Model

grammar-en

Kotoshu

An 85.8 MB int8 ONNX token tagger that detects and corrects grammatical errors fully locally — no data leaves the machine. GECToR architecture (Omelianchuk et al. 2020): distilroberta encoder, 5,000-label closed vocabulary ($KEEP / $REPLACE / $APPEND / $DELETE / $TRANSFORM / $MERGE), iterative apply-and-retag decoding (4 rounds). Scored with the official M2 scorer, noalt gold, --maxunchangedwords 2, 4-round iterative decode: 2.9x LanguageTool; runs in ~6 ms/sentence on a laptop CPU through onnxruntime. labels.json (id -> label map — do not rebuild), vocab.json + merges.txt (distilroberta byte-level BPE). label vocab = top 5,000 from the training mix. real-time (~6 ms/sentence), 85.8 MB int8…

arabic - handwritten-text-recognition - htr - ocr - historical-manuscripts - kraken libraryname: kraken arman-kraken-ocr is a Kraken OCR/HTR model for line-level recognition of handwritten Arabic manuscripts. The model is based on the Muharaf-trained Kraken checkpoint muharafrecbest.mlmodel and was subsequently fine-tuned on the ArMan dataset for historical Arabic manuscript recognition. This checkpoint is a fine-tuned derivative of the Arabic HTR model trained on the Muharaf Corpus: The model was fine-tuned on ArMan as part of our work on historical Arabic manuscript HTR and the AraMS-28k dataset. If you use AraMS-28k in your research, please cite: If you use RefLAM or its annotation…

IdeaLens-LogisticClassifier is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl -o…

IdeaLens-LogisticClassifier-PerItem is an idea-level detector: it judges whose ideas a document contains, not who wrote its words, so a document whose ideas are a person's counts as human however much of its prose an AI wrote. It is one of the detectors released with IdeaLens and trained on the same data. The idealens package (PyPI) runs the whole pipeline: it assigns each document one of the eight formats, extracts the outline with the prompt, role vocabulary and worked examples the detectors were trained with, and scores it with this model and the thresholds in this repo. Input is JSONL with a text field per document. To score outlines you already have, use idealens score outlines.jsonl…

Model

LIBE

Ca' Foscari NLP

basemodel: - sentence-transformers/LaBSE - language-identification - langID - sentence-embedding - similarity-search - low-resource - low-resource-languages

This repository contains the checkpoint for STVG-R1, the first reinforcement learning framework for spatial–temporal video grounding. We finetuned Qwen2.5-VL-7B-Instruct with constructed Visual Prompt Data. If you use STVG-R1 or this checkpoint in your research, please cite

One model, one set of weights, for the IROS 2026 TartanIMU cross-platform inertial odometry challenge. The network predicts body-frame velocity from 6-axis IMU alone for car, drone, quadruped and handheld motion; there are no per-platform experts and no model switching at inference. An all-zero submission scores exactly 1.000. The best public score is ≈ 0.20074, so this The checkpoint is self-contained: it stores the per-channel mean/std used for input standardization plus all architecture hyper-parameters, so inference needs no extra flags or training-time artefacts. --stride 5 and --tta-shifts 0,25,-25 are the inference configuration of the submitted system (see compliance item 3). Using…

This repository is the second of the two single-model checkpoints that team Xunden selected on the Kaggle leaderboard. The first one lives at (RICH2200tv, 31-channel inputs, 200 epochs). The two checkpoints are independent single models: each generated its own selected CSV on its own, and no prediction in either CSV is an ensemble, blend or average of the two. The official per-sequence scoring service was not run on this particular CSV; its recorded result is the Kaggle public leaderboard entry above. The richer companion model in the sister repository reports the official per-sequence service numbers. The checkpoint is self-contained: it stores per-channel mean/std for input…

Model · Feature extraction

TeraAnaly

Tera

Russian speech analysis with one frozen w2v-BERT encoder and independently callable age, gender, emotion, stress and CTC components. Decode audio once, run the encoder once, and reuse its output for any heads you need. Noncommercial use only for the TeraAnaly code and trained heads, under CC-BY-NC-SA-4.0. The original encoder weights retain their upstream CC-BY-4.0 license. See LICENSE.md. Outputs are an integer age in years, the highest gender label, the highest emotion label, and a transcript with + before stressed vowels. Emotion labels are neutral, angry, positive, sad, other; demographic labels are female, male, child (child is an age category inherited from the teacher).…

Sentiment (negative / neutral / positive) for Vietnamese student feedback about university courses. PhoBERT-base fine-tuned on UIT-VSFC with diacritic and teencode augmentation and, at every step, trained to give a real informally typed comment the label a frozen teacher gives its normalized form. Real typing changes 12% of its labels, against 18% for the same model without that training. Shipped as the ONNX FP32 graph the ViFeedback service runs, plus the PyTorch checkpoint it was exported from. Code, evaluation protocol and every number below: https://github.com/ThanhDatVN/ViFeedback-NLP-Service The model expects pyvi word-segmented, lowercase text (giảngviên nhiệttình), at most 96…

Ask anything, get a how-to article. WikiQwen-4B is Qwen3.5-4B fine-tuned to answer every message the same way: as a tidy, step-by-step how-to article in markdown, with a picture caption above each step. Hand those captions to WikiQwen-Illustrator and they become illustrations. Send it a question ("how do I keep basil alive?"), a problem ("my bike chain keeps falling off") or just "hi", and it sends back a guide. Every reply has the same shape: - a # How to … title and a short intro - one or more ## Method N: or ## Part N: sections, or a single ## Steps - an [IMAGE: caption] line above each step, written for the Illustrator to draw - steps written as N. Bold summary. Details…, with bullets…

Ask anything, get a how-to article. WikiQwen-9B is Qwen3.5-9B fine-tuned to answer every message the same way: as a tidy, step-by-step how-to article in markdown, with a picture caption above each step. Hand those captions to WikiQwen-Illustrator and they become illustrations. Send it a question ("how do I keep basil alive?"), a problem ("my bike chain keeps falling off") or just "hi", and it sends back a guide. Every reply has the same shape: - a # How to … title and a short intro - one or more ## Method N: or ## Part N: sections, or a single ## Steps - an [IMAGE: caption] line above each step, written for the Illustrator to draw - steps written as N. Bold summary. Details…, with bullets…

Give it a caption, get a step illustration. WikiQwen-Illustrator is a style LoRA for [FLUX.2 [klein] base 4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B). It turns a one-to-three-sentence scene description into a clean, friendly how-to illustration at 512x384. It is the picture half of WikiQwen: the chat models (4B, 9B, 35B-A3B) write an [IMAGE: caption] line above every step, and this LoRA draws it. Inside a WikiQwen article, a step looks like this. The first line is the Illustrator's prompt, and the second is the step it illustrates: - No trigger word. The training captions describe content only, with no style words, so the style comes along automatically. It takes a…