Villanova-2B-VL-2603 is a fully open, multilingual Vision-Language Model developed by Villanova.AI. Part of the Villanova project, it extends our text-only Villanova-2B-2603 to visual understanding while preserving native support for five European languages. All model weights, training data sources, and training details are publicly released.
Built on a LLaVA-style architecture pairing a SigLIP vision encoder with the Villanova-2B-Base-2603 language backbone, this ~2.8B-parameter model delivers strong multimodal understanding, visual question answering, and multilingual image captioning under a fully open Apache 2.0 license.
European-focused, fully open VLM released under Apache 2.0
Native multilingual support for 5 European languages: English, French, German, Italian, and Spanish, including multilingual image captioning (XM3600) and visual instruction following
Broad visual understanding across general VQA (RealWorldQA, CVQA, MME) and multilingual benchmarks (Multi-MMBench, Multi-AI2D)
Preserves text-only capabilities of the Villanova-2B-2603 language backbone through text-only data mixing in Stage 2
Only ~2.8B parameters, efficient enough for single-GPU inference
Projector-only alignment on multilingual image-caption pairs
Stage 2
LLM unfrozen, vision tower frozen, visual instruction tuning on a fullmix recipe (~1.08M samples)
Languages
English, French, German, Italian, Spanish
Max Sequence Length
32,768 tokens
Precision
bfloat16
License
Apache 2.0
Training Recipe (Stage 1: Projector Alignment)
Stage 1 aligns the vision encoder output to the language model embedding space by training only the multimodal projector, with both the vision tower and the LLM fully frozen. This is a lightweight warmup that teaches the projector how to map SigLIP visual features into the Villanova-2B token space before any instruction tuning.
Both stages use only permissively-licensed data (no GPT/Claude-generated content). The curated multilingual derivatives (the Multi-* datasets, translated and post-processed in EN/DE/ES/FR/IT) are released by Villanova.AI on the HuggingFace Hub.
1import torch
2from PIL import Image
3from transformers import AutoProcessor, LlavaForConditionalGeneration
45model_name ="VillanovaAI/Villanova-2B-VL-2603"6device ="cuda"78processor = AutoProcessor.from_pretrained(model_name)9model = LlavaForConditionalGeneration.from_pretrained(10 model_name,11 dtype=torch.bfloat16,12).to(device)13model.eval()1415image = Image.open("example.jpg").convert("RGB")1617# The `<image>` placeholder inside the content string marks where the18# image tokens will be inserted by the processor.19messages =[20{"role":"user","content":"<image>\nDescribe this image in detail."},21]2223prompt = processor.apply_chat_template(messages, add_generation_prompt=True)24inputs = processor(images=image, text=prompt, return_tensors="pt").to(device, torch.bfloat16)2526with torch.no_grad():27 output_ids = model.generate(**inputs, max_new_tokens=256, do_sample=False)2829response = processor.decode(output_ids[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)30print(response)
Evaluation
Villanova-2B-VL-2603 was evaluated using VLMEvalKit on a suite of standard and multilingual VLM benchmarks covering multiple-choice reasoning, general visual question answering, hallucination robustness, and cross-lingual visual understanding. All evaluations use exact_matching judging (no LLM-as-judge) for full reproducibility.
We compare against Salamandra-VL-7B, a strong European VLM built on the same language family.
Despite using less than a third of the parameters (~2.8B vs ~8.9B), Villanova-2B-VL-2603 matches Salamandra-VL-7B overall, with particular strengths on general VQA and multilingual benchmarks. Multilingual benchmarks are reported as the average across EN/DE/ES/FR/IT.
Category
Benchmark
Salamandra-VL-7B
Villanova-2B-VL-2603
MCQ / Reasoning
MMBench
51.9
52.6
MCQ / Reasoning
MMStar
38.9
32.7
MCQ / Reasoning
AI2D
58.5
51.1
MCQ / Reasoning
ScienceQA
65.0
62.4
General VQA
RealWorldQA
45.5
46.8
General VQA
CVQA
32.9
37.6
General VQA
MME
1369
1565
Hallucination
POPE
86.8
81.1
OCR / Document
OCRBench
558
377
Multilingual (avg)
Multi-MMBench
59.7
61.0
Multilingual (avg)
Multi-AI2D
58.4
66.4
Multilingual (avg)
Multi-MMStar
50.2
47.6
Overall
Average (0-100 benchmarks)
54.8
53.9
The Overall row is the unweighted average across the 10 benchmarks on the 0-100 scale. MME and OCRBench are excluded because they use different scoring scales (0-2800 and 0-1000 respectively).
Multilingual Evaluation (Per-Language Detail)
The multilingual benchmarks (Multi-MMBench, Multi-AI2D, Multi-MMStar) are extensions of the standard benchmarks with parallel test sets in 5 European languages. Below is the per-language breakdown.
Benchmark
Model
DE
EN
ES
FR
IT
Avg
Multi-MMBench
Salamandra-VL-7B
59.3
64.8
62.7
57.0
54.7
59.7
Multi-MMBench
Villanova-2B-VL-2603
60.8
62.8
58.9
60.6
61.6
61.0
Multi-AI2D
Salamandra-VL-7B
57.0
67.1
62.2
53.6
52.2
58.4
Multi-AI2D
Villanova-2B-VL-2603
66.6
68.1
65.1
66.9
65.4
66.4
Multi-MMStar
Salamandra-VL-7B
46.6
56.5
52.3
48.2
47.3
50.2
Multi-MMStar
Villanova-2B-VL-2603
45.9
50.1
47.0
49.2
45.7
47.6
Key takeaways:
Competitive overall average (53.9 vs 54.8) against a model with ~3.2x more parameters
Wins on general VQA: RealWorldQA, CVQA, and MME all outperform Salamandra-VL-7B
Solid multilingual capability across EN/DE/ES/FR/IT, with a particularly strong Multi-AI2D improvement (+8.0 avg, wins on all 5 languages) over Salamandra-VL-7B
Balanced per-language performance: on Multi-AI2D and Multi-MMBench, Villanova performs uniformly across DE/EN/ES/FR/IT (no language collapse)
Intended Use
Multilingual image captioning and description
Visual question answering (single-image)
Document and chart understanding (OCR-light tasks)
Multimodal instruction following in EN/DE/ES/FR/IT
Research on fully-open European VLMs
Limitations
Single-image inference only (no multi-image or video support)
OCR quality on dense, small-text documents is limited compared to specialized OCR-heavy VLMs
As with all VLMs, outputs can contain hallucinations; users should verify factual claims
License
This model is released under the Apache 2.0 License. The training data used for Stage 2 was selected to allow permissive commercial use (no GPT/Claude-generated content).
