Quick start
import torchfrom transformers import AutoModelForImageTextToText, AutoProcessor, AutoTokenizer model_id = "VillanovaAI/Villanova2-15B-A3B-VL" # tokenizer_plain/ keeps the training tokenizer in every transformers version (see the warning above)tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder="tokenizer_plain")processor = AutoProcessor.from_pretrained(model_id, tokenizer=tokenizer)model = AutoModelForImageTextToText.from_pretrained( model_id, dtype=torch.bfloat16, device_map="auto",) messages = [ { "role": "user", "content": [ {"type": "image", "path": "invoice.png"}, {"type": "text", "text": "Estrai numero, data e totale della fattura in formato JSON."}, ], }]inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device) output = model.generate(**inputs, max_new_tokens=1024, do_sample=False)print(processor.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
With vLLM, point --tokenizer to the tokenizer_plain/ folder of a local copy:
hf download VillanovaAI/Villanova2-15B-A3B-VL --local-dir Villanova2-15B-A3B-VLvllm serve ./Villanova2-15B-A3B-VL --tokenizer ./Villanova2-15B-A3B-VL/tokenizer_plain \ --max-model-len 32768 --limit-mm-per-prompt '{"image": 8}'
The tested stacks are transformers 5.17 as is, transformers 5.19 and 5.20 with tokenizer_plain/, and vLLM 0.32. transformers 4.x does not support the Qwen3.5-MoE architecture. All evaluations below use greedy decoding, the default of the shipped generation_config.json. When no system message is given, the chat template adds a default one ("You are Villanova, a helpful AI assistant built by Villanova.AI.").
The shipped processor encodes every image with 2,304 to 3,328 visual tokens (one token per 32 × 32 pixels after the 2 × 2 merge; smaller images are upscaled), and this is the setting used for all the numbers in this card.
The language model is hybrid: only 7 of its 28 layers hold a KV cache, the other 21 are Gated DeltaNet recurrent layers with a fixed-size state. Long multi-page prompts cost far less memory than the parameter count suggests.
Quantized version
Villanova2-15B-A3B-VL-GGUF holds the same weights quantized for llama.cpp, for edge inference; llama cli handles images natively and downloads the projector automatically:
llama cli -hf VillanovaAI/Villanova2-15B-A3B-VL-GGUF:Q8_0
Intended use
Table with columns: Good fit, Not a good fit without extra work| Good fit | Not a good fit without extra work |
|---|
| Document transcription to text or Markdown, single and multi-page | Safety-critical or high-stakes decisions without human review |
| Key-value and named-entity extraction from business documents | Scene-text reading in natural photos (TextVQA 61.7) |
| Page and document summarization and translation in EN · IT · DE · ES · FR | Video understanding |
| Question answering over pages, tables and charts | Languages outside the five core ones, where quality is not guaranteed |
| Fine-tuning into domain-specific document assistants | Use with transformers ≥ 5.18 without the tokenizer_plain/ workaround |
The family
Both vision-language models start from the matching Villanova2 base model and see the same multimodal training data in the same three main stages; the 2B adds a short high-resolution stage at the end.
Architecture
Table with columns: Component, Value| Component | Value |
|---|
| Language model | Villanova2-15B-A3B-Base: Qwen3-Next hybrid decoder, 28 layers (21 Gated DeltaNet + 7 full attention), hidden 2,560 |
| Experts | 64 routed SwiGLU experts, top-8 per token, + 1 shared expert |
| Vision encoder | Qwen3.5 ViT: 24 layers, hidden 1,024, 16 heads, patch 16, native dynamic resolution |
| Projector | 2 × 2 spatial merge into the 2,560-dim LM space |
| Initialization | LM from Villanova2-15B-A3B-Base; vision encoder and projector from Qwen3.5-4B |
| Vocabulary | 130,822-token Villanova BPE tokenizer · 130,822 embedding rows, untied input/output |
| Positional encoding | Multimodal RoPE (MRoPE) over text, height and width, partial (25% of the head dimension) |
| Language model parameters | 15,496,439,744 (3,935,889,344 active per token) |
| Vision encoder parameters | 306,242,560 |
| Projector parameters | 27,271,680 |
| Total parameters | 15,829,953,984 |
The model uses only native transformers classes: no trust_remote_code, and the same checkpoint loads in transformers and vLLM.
Training
Three stages, one epoch each
Table with columns: Stage, What is trained, Examples, Unique images, Tokens, Learning rate (LM / projector / ViT)| Stage | What is trained | Examples | Unique images | Tokens | Learning rate (LM / projector / ViT) |
|---|
| 1 · Alignment | Projector only (LM and ViT frozen) | 1,046,387 | 611,969 | 1.70B | — / 1 × 10⁻³ / — |
| 1.5 · Knowledge | Full model | 3,194,343 | 3,179,728 | 6.61B | 5 × 10⁻⁶ / 1 × 10⁻⁵ / 1 × 10⁻⁶ |
| 2 · Instruction tuning | Full model | 8,026,641 | 4,896,956 | 12.19B | 5 × 10⁻⁶ / 1 × 10⁻⁵ / 1 × 10⁻⁶ |
| Total | | 12,267,371 | 7,360,035 | 20.50B | |
Token counts include the visual tokens. Stages 1.5 and 2 pack examples into sequences of up to 16,384 tokens; the 240 examples longer than that (0.002%) are truncated to that length, and the 9 of them whose cut would fall inside an image are dropped. Unique images are counted across all stages: an image reused in a later stage counts once.
Data
[!IMPORTANT]
We used synthetic data only from models released under a permissive licence. Where the origin of a dataset was explicit, we removed the proprietary parts, tracing provenance as carefully as we could; where the origin was doubtful, we did not use the data.
This distinguishes Villanova2 from many open models, which are post-trained on popular datasets containing answers written by closed models.
Every source is cleared for commercial use. Sources marked (ML) were translated by us from English into Italian, German, Spanish and French; Villanova DU marks the document data we built from FinePDFs pages in the five languages. Where Villanova DU and CommonCanvas needed generated text (questions, answers, summaries, captions), it was produced with Qwen3.5-122B-A10B, self-hosted. Every training example whose image is byte-identical to a page of the MDU test set is removed.
Stage 1 · Alignment · 1.05M examples
Table with columns: Source family, Sources, Role| Source family | Sources | Role |
|---|
| Captions | PixMo-Cap, brief and detailed (ML) | Align the vision features with the language model |
Stage 1.5 · Knowledge · 3.19M examples
Table with columns: Source family, Sources, Role| Source family | Sources | Role |
|---|
| Captions | PixMo-Cap (ML), CommonCanvas captions | Dense image descriptions |
| Document transcription | Villanova DU: page to Markdown and to text, single and multi-page | Read long documents end to end |
| Encyclopedic pages | Nemotron-VLM-Dataset-v2 wiki pages (10 languages) | Rendered pages transcribed to Markdown |
| Charts | PlotQA (ML) | Chart question answering |
Stage 2 · Instruction tuning · 8.03M examples
Table with columns: Source family, Sources, Role| Source family | Sources | Role |
|---|
| General VQA | VQAv2, GQA, A-OKVQA, VizWiz (from FineVision), PixMo-AskModelAnything (ML), PixMo-Cap-QA (ML), CommonCanvas VQA, Crossmodal-3600 | Natural-image questions and captions |
| Scene text and OCR | TextVQA, DocVQA (from FineVision), TextCaps, SynthDoG-en | Reading text in images |
| Documents | Villanova DU: QA, key-value and entity extraction, summarization, translation, classification, layout, anonymization, single and multi-page; Docmatix; FinQA (ML) | Document understanding in the five languages |
| Tables, charts, formulas | PubTabNet, SynthTabNet, SynthChartNet, SynthFormulaNet, CoSyn-400K (ML), FigureQA (ML), PlotQA (ML), Nemotron SparseTables | Structure recognition and chart QA |
| Science, math, reasoning | AI2D, CLEVR and CLEVR-Math (ML), Nemotron-VLM-Dataset-v2 reasoning subsets (ChartQA, DocVQA, GeomVerse, Mulberry, PlotQA, Visual7W, VisualWebInstruct) | Diagrams and step-by-step answers |
| Grounding | Nemotron-VLM-Dataset-v2 Open Images boxes | Object localization |
| Encyclopedic pages | Nemotron-VLM-Dataset-v2 wiki pages (10 languages) | Long-page transcription |
| Text only | The Instruct set of the Villanova2-15B-A3B SFT mixture, including Multi-SFT (Baiamonte et al., 2026), 2.03M conversations in the five languages | Keep the language model's chat and reasoning skills |
Evaluation
Every model in the tables, ours and the external ones, goes through the same pipeline with greedy decoding. For the VLMEvalKit and multilingual tables: VLMEvalKit data, splits and metrics, up to 2,048 new tokens, the LLaVA-1.5 / lmms-eval answer-format instructions, and rule-based answer extraction (exact matching, no LLM judge). Scores are raw. Because the protocol is the same for every model, numbers can differ from those published in other model cards. Openness: weight = open weights only; weight & data = open weights plus documented training data. In the comparison tables, bold marks the best score in each column and underline the best among the open weight & data EU models.
The comparison covers the strongest open models up to about 16B dense or 30B mixture-of-experts parameters, every EU vision-language model that completed the same protocol, from 4.3B to 14B, and the two Apertus models (8B and 70B). EuroVLM-9B is not reported: its run did not produce usable outputs.
General vision-language benchmarks (VLMEvalKit)
Table with columns: Model, Openness, Origin, Params, Avg, AI2D, ChartQA, DocVQA, MMBench, MMMU, MMStar, MTVQA, OCRBench, POPE, RWQA, SEED, SQA, TextVQA| Model | Openness | Origin | Params | Avg | AI2D | ChartQA | DocVQA | MMBench | MMMU | MMStar | MTVQA | OCRBench | POPE | RWQA | SEED | SQA | TextVQA |
|---|
| Qwen3-VL-30B-A3B-Instruct | weight | non-EU | 31.1B / 3.3B | 78.0 | 82.5 | 85.4 | 95.2 | 85.7 | 53.2 | 66.1 | 33.8 | 90.2 | 89.2 | 74.8 | 78.0 | 95.8 | 84.6 |
| Qwen3-VL-8B-Instruct | weight | non-EU | 8.8B | 76.8 | 84.0 | 83.0 | 95.7 | 84.1 | 48.7 | 64.4 | 32.5 | 89.7 | 87.7 | 72.7 | 77.5 | 94.7 | 83.1 |
| InternVL3.5-14B | weight | non-EU | 15.1B | 76.5 | 85.2 | 86.5 | 93.0 | 83.4 | 59.0 | 68.9 | 27.8 | 82.8 | 86.5 | 70.2 | 77.3 | 96.5 | 77.8 |
| Villanova2-15B-A3B-VL | weight & data | EU | 15.8B / 4.3B | 70.0 | 79.6 | 76.3 | 89.0 | 77.6 | 50.7 | 55.7 | 23.2 | 74.2 | 88.8 | 70.8 | 77.0 | 85.8 | 61.7 |
| Ministral-3-14B-Instruct | weight | EU | 14.0B | 69.5 | 79.2 | 77.4 | 92.2 | 74.3 | 47.8 | 51.9 | 28.3 | 74.8 | 81.2 | 64.6 | 71.2 | 86.7 | 74.4 |
| Pixtral-12B | weight | EU | 12.7B | 68.7 | 79.8 | 72.2 | 88.9 | 76.8 | 46.1 | 52.6 | 20.2 | 68.2 | 85.5 | 66.1 | 72.1 | 88.5 | 75.7 |
| gemma-4-26B-A4B | weight | non-EU | 25.8B / 3.8B | 67.2 | 71.1 | 33.0 | 87.7 | 82.5 | 43.3 | 53.2 | 33.6 | 83.9 | 84.5 | 69.3 | 77.7 | 85.3 | 68.3 |
| Ministral-3-8B-Instruct | weight | EU | 8.9B | 66.9 | 77.3 | 68.0 | 91.2 | 71.0 | 40.3 | 47.9 | 28.9 | 74.7 | 81.5 | 61.2 | 69.2 | 85.4 | 73.1 |
| gemma-3-12b-it | weight | non-EU | 12.2B | 64.9 | 81.2 | 37.6 | 80.5 | 73.8 | 44.7 | 54.7 | 26.3 | 75.7 | 84.8 | 61.6 | 71.9 | 83.0 | 67.7 |
| Apertus-1.5-70B | weight & data | non-EU (CH) | 72.0B | 63.1 | 74.3 | 62.2 | 82.3 | 63.2 | 41.3 | 46.3 | 21.7 | 64.9 | 85.9 | 60.9 | 73.0 | 81.6 | 62.3 |
| Apertus-1.5-8B | weight & data | non-EU (CH) | 8.9B | 63.0 | 73.6 | 72.4 | 83.9 | 57.8 | 38.7 | 45.3 | 20.9 | 65.8 | 83.6 | 58.0 | 72.3 | 83.8 | 62.4 |
| Idefics3-8B | weight | non-EU | 8.5B | 58.7 | 75.8 | 31.6 | 70.8 | 73.8 | 42.0 | 53.6 | 9.5 | 44.4 | 84.2 | 62.9 | 72.8 | 90.0 | 51.7 |
| TowerVision-9B ‡ | weight & data | EU | 9.7B | 58.4 | 67.6 | 26.4 | 70.5 | 71.4 | 29.4 | 44.1 | 21.0 | 68.5 | 87.9 | 65.0 | 74.1 | 76.3 | 57.0 |
| Velvet2-3B-Vision | weight | EU | 4.3B | 58.0 | 68.7 | 54.9 | 66.4 | 66.2 | 34.0 | 41.1 | 7.9 | 58.8 | 88.2 | 57.8 | 66.5 | 76.3 | 67.8 |
‡ TowerVision ends most short answers with a period, which exact matching counts as wrong; ChartQA and TextVQA are the most affected. It runs on transformers 4.57, the only stack that passes it the images.
Villanova2-15B-A3B-VL places 4th of 14 and first among the EU models, ahead of Ministral-3-14B (+0.5), Pixtral-12B (+1.3), Ministral-3-8B (+3.1), Apertus-1.5-70B (+6.9, four times larger), Apertus-1.5-8B (+7.0), Idefics3-8B (+11.3), TowerVision-9B (+11.6), Velvet2-3B-Vision (+12.0) and both Gemma models. It is close to the leaders on perception and real-world questions (POPE 88.8, SEED 77.0, RealWorldQA 70.8) and ahead of Qwen3-VL-8B on MMMU (50.7). The 6.5 to 8.0-point gap to Qwen3-VL-8B, Qwen3-VL-30B-A3B and InternVL3.5-14B is widest on scene text and OCR (TextVQA, OCRBench, MTVQA), charts and MMStar.
Multilingual benchmarks
Seven VLMEvalKit benchmarks translated into German, Spanish, French and Italian, plus the English originals: 35 tasks. Each prompt, including the answer-format instruction, is in the language of the example. Columns are 5-language averages; MME is the perception + cognition score (0–2,800), rescaled to 0–100 in the overall average.
Table with columns: Model, Openness, Origin, Params, Avg, AI2D, MMBench, MME, POPE, RealWorldQA, ScienceQA, VQAv2| Model | Openness | Origin | Params | Avg | AI2D | MMBench | MME | POPE | RealWorldQA | ScienceQA | VQAv2 |
|---|
| Qwen3-VL-30B-A3B-Instruct | weight | non-EU | 31.1B / 3.3B | 82.1 | 79.5 | 89.9 | 2405 | 87.3 | 68.5 | 93.7 | 69.9 |
| Qwen3-VL-8B-Instruct | weight | non-EU | 8.8B | 80.8 | 80.5 | 88.9 | 2317 | 86.3 | 66.2 | 91.9 | 68.9 |
| InternVL3.5-14B | weight | non-EU | 15.1B | 79.8 | 81.0 | 87.3 | 2290 | 87.0 | 63.6 | 94.3 | 63.5 |
| Villanova2-15B-A3B-VL | weight & data | EU | 15.8B / 4.3B | 77.3 | 76.4 | 84.8 | 2074 | 86.4 | 62.7 | 84.5 | 72.0 |
| Pixtral-12B | weight | EU | 12.7B | 74.9 | 72.0 | 83.2 | 1985 | 85.4 | 57.4 | 83.3 | 71.8 |
| gemma-4-26B-A4B | weight | non-EU | 25.8B / 3.8B | 74.6 | 76.1 | 88.7 | 2224 | 83.3 | 62.8 | 85.5 | 46.6 |
| gemma-3-12b-it | weight | non-EU | 12.2B | 72.1 | 77.1 | 83.8 | 2111 | 84.5 | 46.0 | 82.2 | 55.5 |
| Ministral-3-14B-Instruct | weight | EU | 14.0B | 71.1 | 75.0 | 83.4 | 1850 | 80.4 | 55.4 | 85.8 | 51.9 |
| Ministral-3-8B-Instruct | weight | EU | 8.9B | 70.7 | 71.2 | 81.6 | 1786 | 80.0 | 56.6 | 83.3 | 58.4 |
| Apertus-1.5-8B | weight & data | non-EU (CH) | 8.9B | 69.2 | 69.6 | 71.8 | 1611 | 84.3 | 56.1 | 80.9 | 64.4 |
| Apertus-1.5-70B | weight & data | non-EU (CH) | 72.0B | 68.7 | 69.7 | 73.6 | 1552 | 85.8 | 57.0 | 79.6 | 59.9 |
| Velvet2-3B-Vision | weight | EU | 4.3B | 67.2 | 63.5 | 74.5 | 1655 | 87.2 | 49.3 | 72.0 | 64.5 |
| TowerVision-9B ‡ | weight & data | EU | 9.7B | 66.4 | 66.9 | 80.0 | 1675 | 78.5 | 50.5 | 75.4 | 53.9 |
| Idefics3-8B | weight | non-EU | 8.5B | 66.2 | 69.6 | 80.5 | 1868 | 83.7 | 46.3 | 86.2 | 30.2 |
Again 4th of 14, and 2.4 points above the best EU model, Pixtral-12B. On open-ended answers the model is the strongest in the table: its 5-language VQAv2 average (72.0) is the highest of the 14, and 25 points above gemma-4-26B-A4B.
Per-language profile of Villanova2-15B-A3B-VL.
Table with columns: Language, Avg, AI2D, MMBench, MME, POPE, RealWorldQA, ScienceQA, VQAv2| Language | Avg | AI2D | MMBench | MME | POPE | RealWorldQA | ScienceQA | VQAv2 |
|---|
| English | 79.2 | 77.8 | 84.8 | 2153 | 87.4 | 63.9 | 86.2 | 77.1 |
| Italian | 75.5 | 76.8 | 84.9 | 2023 | 87.2 | 55.6 | 83.8 | 67.8 |
| German | 76.8 | 75.0 | 84.0 | 2057 | 82.5 | 66.1 | 84.2 | 72.5 |
| Spanish | 77.0 | 76.7 | 85.2 | 2108 | 88.2 | 59.7 | 84.0 | 70.1 |
| French | 77.9 | 75.9 | 85.0 | 2027 | 86.5 | 68.2 | 84.4 | 72.7 |
The five languages sit within 3.7 points of each other, and MMBench is flat across them (84.0–85.2). Italian is the language with the largest gap from English, driven by RealWorldQA and VQAv2.
Document understanding (MDU)
MDU is our multilingual document-understanding benchmark: real-world PDF documents in the five languages, 18 task groups, each scored 0–100 with a reference-based metric (character-level similarity for transcription, micro-F1 on the extracted values for key-value and entity extraction, chrF / chrF++ for summaries and translations, numeric match with tolerance for table and figure questions, token recall for open questions, constraint satisfaction for conditional summaries; no LLM judge). Generation is greedy, with a 65,536-token context. The table groups the 18 task groups into five families; MDU-18 is their plain average. The Apertus-1.5-70B has no MDU run yet.
Table with columns: Model, Openness, Origin, Params, MDU-18, Transcr., QA, Extr., Summ., Transl.| Model | Openness | Origin | Params | MDU-18 | Transcr. | QA | Extr. | Summ. | Transl. |
|---|
| Villanova2-15B-A3B-VL | weight & data | EU | 15.8B / 4.3B | 67.0 | 74.2 | 60.6 | 63.9 | 65.4 | 73.5 |
| Qwen3-VL-8B-Instruct | weight | non-EU | 8.8B | 62.8 | 70.1 | 65.6 | 43.6 | 63.6 | 54.6 |
| Ministral-3-14B-Instruct | weight | EU | 14.0B | 60.5 | 63.3 | 63.1 | 43.7 | 61.3 | 61.0 |
| Qwen3-VL-30B-A3B-Instruct | weight | non-EU | 31.1B / 3.3B | 60.1 | 72.6 | 60.6 | 35.7 | 59.5 | 52.3 |
| Ministral-3-8B-Instruct | weight | EU | 8.9B | 59.1 | 60.5 | 62.6 | 41.4 | 62.4 | 57.7 |
| InternVL3.5-14B | weight | non-EU | 15.1B | 57.6 | 67.0 | 58.1 | 38.1 | 61.3 | 46.3 |
| gemma-4-26B-A4B | weight | non-EU | 25.8B / 3.8B | 56.0 | 55.1 | 57.3 | 45.3 | 63.4 | 54.2 |
| Pixtral-12B | weight | EU | 12.7B | 51.3 | 60.8 | 51.2 | 28.3 | 56.7 | 42.8 |
| gemma-3-12b-it | weight | non-EU | 12.2B | 50.6 | 50.8 | 51.9 | 36.2 | 57.9 | 49.2 |
| Apertus-1.5-8B | weight & data | non-EU (CH) | 8.9B | 34.7 | 45.9 | 23.8 | 17.4 | 53.5 | 28.9 |
| Idefics3-8B | weight | non-EU | 8.5B | 31.7 | 30.5 | 37.6 | 18.2 | 42.4 | 15.0 |
| Velvet2-3B-Vision | weight | EU | 4.3B | 29.9 | 33.8 | 31.8 | 7.0 | 41.6 | 19.6 |
| TowerVision-9B †§ | weight & data | EU | 9.7B | 25.0 | 19.7 | 26.6 | 20.1 | 42.2 | 12.3 |
| salamandra-7b-vision † | weight & data | EU | 8.2B | 14.0 | 10.0 | 16.9 | 4.4 | 27.8 | 3.8 |
† The harness flags these runs as not comparable: too many answers were cut by the generation limit (30.2% for TowerVision-9B, 22.5% for salamandra-7b-vision).
§ TowerVision has an 8,192-token context: longer prompts are scored as empty answers, which puts Doc2Markdown, Doc2Text and Document Translation at 0.
Table with columns: Task group, Villanova2-15B-A3B-VL, Qwen3-VL-8B-Instruct, Ministral-3-14B-Instruct, Qwen3-VL-30B-A3B-Instruct, Ministral-3-8B-Instruct, InternVL3.5-14B, gemma-4-26B-A4B, Pixtral-12B, gemma-3-12b-it, Apertus-1.5-8B, Idefics3-8B, Velvet2-3B-Vision, TowerVision-9B †§, salamandra-7b-vision †| Task group | Villanova2-15B-A3B-VL | Qwen3-VL-8B-Instruct | Ministral-3-14B-Instruct | Qwen3-VL-30B-A3B-Instruct | Ministral-3-8B-Instruct | InternVL3.5-14B | gemma-4-26B-A4B | Pixtral-12B | gemma-3-12b-it | Apertus-1.5-8B | Idefics3-8B | Velvet2-3B-Vision | TowerVision-9B †§ | salamandra-7b-vision † |
|---|
| Transcription | | | | | | | | | | | | | | |
| Doc2Markdown | 68.9 | 76.8 | 73.2 | 77.4 | 66.5 | 74.8 | 58.9 | 66.4 | 52.0 | 32.4 | 26.4 | 15.0 | 0.0 | 0.0 |
| Doc2Text | 78.4 | 64.5 | 58.8 | 75.4 | 52.8 | 65.2 | 50.8 | 61.0 | 54.2 | 31.7 | 18.0 | 13.9 | 0.0 | 0.0 |
| Image2Markdown | 89.8 | 82.9 | 71.4 | 84.2 | 70.7 | 76.7 | 60.0 | 72.0 | 41.2 | 67.5 | 42.6 | 56.4 | 35.7 | 11.1 |
| Image2Text | 93.4 | 75.1 | 67.0 | 80.5 | 66.0 | 73.7 | 63.8 | 68.0 | 62.6 | 67.6 | 33.3 | 54.8 | 25.2 | 10.8 |
| Passage Extraction | 40.3 | 51.1 | 46.2 | 45.7 | 46.5 | 44.7 | 42.1 | 36.4 | 44.0 | 30.3 | 32.2 | 28.6 | 37.9 | 28.2 |
| QA | | | | | | | | | | | | | | |
| Question Answering | 65.4 | 64.8 | 61.7 | 57.1 | 61.8 | 60.9 | 52.2 | 54.9 | 53.2 | 29.5 | 48.5 | 40.2 | 51.0 | 39.6 |
| TableQA | 66.9 | 73.2 | 74.3 | 64.7 | 74.1 | 64.6 | 61.6 | 55.7 | 59.2 | 33.4 | 32.8 | 31.6 | 45.9 | 29.0 |
| FigureQA | 55.5 | 62.4 | 60.5 | 58.3 | 60.3 | 51.9 | 55.4 | 46.4 | 52.5 | 29.2 | 35.3 | 36.4 | 43.6 | 32.7 |
| Multi-page QA | 61.4 | 63.6 | 58.7 | 59.9 | 59.7 | 60.1 | 58.6 | 52.2 | 40.0 | 18.7 | 45.9 | 36.4 | 10.3 | 0.1 |
| Multi-page TableQA | 61.3 | 65.2 | 67.4 | 62.4 | 64.6 | 60.1 | 59.3 | 54.3 | 56.0 | 20.6 | 31.0 | 20.4 | 5.2 | 0.0 |
| Multi-page figureQA | 52.9 | 64.1 | 55.8 | 61.1 | 55.1 | 51.0 | 56.5 | 43.9 | 50.7 | 11.7 | 31.8 | 25.9 | 3.6 | 0.2 |
| Extraction | | | | | | | | | | | | | | |
| Key-value Extraction | 60.2 | 46.2 | 43.8 | 28.1 | 42.1 | 36.9 | 46.2 | 28.6 | 33.2 | 16.4 | 17.0 | 8.8 | 18.6 | 4.8 |
| Named-entities extraction | 67.5 | 41.1 | 43.6 | 43.3 | 40.6 | 39.2 | 44.4 | 28.0 | 39.2 | 18.3 | 19.4 | 5.2 | 21.6 | 4.0 |
| Summarization | | | | | | | | | | | | | | |
| Conditional Summarization | 84.0 | 87.8 | 83.9 | 83.6 | 85.1 | 85.6 | 90.4 | 77.2 | 90.9 | 85.9 | 71.6 | 64.0 | 80.9 | 65.0 |
| Document Summarization | 52.6 | 49.4 | 48.0 | 43.1 | 49.1 | 47.2 | 50.3 | 44.3 | 38.5 | 30.0 | 25.4 | 23.0 | 5.9 | 0.0 |
| Page Summarization | 59.6 | 53.6 | 52.1 | 51.8 | 53.1 | 51.1 | 49.6 | 48.7 | 44.5 | 44.4 | 30.1 | 37.8 | 39.7 | 18.4 |
| Translation | | | | | | | | | | | | | | |
| Document Translation | 69.5 | 47.7 | 57.9 | 44.6 | 52.9 | 38.2 | 51.8 | 39.6 | 46.8 | 21.6 | 10.1 | 13.0 | 0.0 | 0.0 |
| Page Translation | 77.6 | 61.5 | 64.0 | 60.1 | 62.6 | 54.5 | 56.6 | 45.9 | 51.7 | 36.2 | 19.8 | 26.3 | 24.6 | 7.6 |
On documents, Villanova2-15B-A3B-VL is first of the 14 models: 4.2 points above Qwen3-VL-8B, 6.5 above Ministral-3-14B, 6.9 above Qwen3-VL-30B-A3B (a model twice its size) and 15.7 above Pixtral-12B. The lead is largest where the answer has to be written in the document's language: extraction (+18.6 over the runner-up, key-value 60.2 and named entities 67.5) and translation (+12.5), then summarization (+1.8) and transcription (+1.6). Table and figure QA is the one family where it trails, 5.0 points behind Qwen3-VL-8B.
The general and multilingual benchmarks are the complementary side: the model leads every EU system in the table, from 4.3B to 14B, and is close to the leaders on perception, but behind the Qwen and InternVL 8B–30B models on scene text, OCR and charts. It is the strongest model here for multilingual document work and the strongest EU one overall.
Fine-tuning
- Start from this checkpoint for document or domain adaptation; LoRA on the language model is enough for most extraction and formatting tasks.
- Load the tokenizer from
tokenizer_plain/ (or use transformers 5.17) for training as well as inference: fine-tuning with the wrong tokenizer would teach the model a different segmentation of numbers and whitespace.
- Keep the shipped chat template: it adds the BOS token and the
<|im_start|> / <|im_end|> turn markers the model was trained with, and generation stops on <|im_end|>.
- The model was trained in BF16 with DeepSpeed ZeRO-3 on packed sequences of up to 16,384 tokens; BF16 is the recommended default for fine-tuning.
- Train at the resolution you will serve at. Dense pages benefit from the shipped 2,304-token minimum.
License, citation, acknowledgements
License. Apache 2.0. The vision encoder and projector are initialized from Qwen3.5-4B, released under Apache 2.0; the language model is Villanova2-15B-A3B-Base.
Citation. If you use Villanova2-15B-A3B-VL, please cite the technical report and the inaugural Villanova paper, Multi-SFT for the multilingual data, and the multilingual vision-language training and evaluation resources:
@inproceedings{simonazzi2026villanova2b, author = {Simonazzi, Stefano and Zeinalipour, Kamyar and Gabburo, Matteo and Baiamonte, Daniela and Fano, Elena and Zugarini, Andrea}, title = {Villanova-2B: European Open Multilingual and Multimodal Language Models}, booktitle = {Proceedings of the Twelfth Italian Conference on Computational Linguistics (CLiC-it 2026)}, editor = {Basile, Valerio and Croce, Danilo and Passaro, Lucia and Pirrone, Roberto}, year = {2026}, note = {In press}}
@techreport{villanova2_2026, title = {Villanova2 --- Technical Report}, author = {{Villanova.AI}}, institution = {Villanova.AI}, year = {2026}, note = {Coming soon}}
@misc{baiamonte2026multisft, author = {Baiamonte, Daniela and Zeinalipour, Kamyar and Fano, Elena and Simonazzi, Stefano and Zugarini, Andrea}, title = {Multi-SFT: Permissively-Licensed Multilingual Instruction-Following}, year = {2026}, eprint = {2610.22086}, archivePrefix = {arXiv}, url = {https://arxiv.org/abs/2610.22086}}
@article{baiamonte2026multilingual, title={Multilingual Training and Evaluation Resources for Vision-Language Models}, author={Baiamonte, Daniela and Fano, Elena and Gabburo, Matteo and Simonazzi, Stefano and Rigutini, Leonardo and Zugarini, Andrea}, journal={arXiv preprint arXiv:2604.18347}, year={2026}}
Acknowledgements. This work is part of the Villanova project, financed by IPCEI-CIS, Prog. n. SA. 102519 — CUP B29J24000850005.