LFM2.5-VL-3B extends vision-language capabilities with four major improvements
The new model from LiquidAI brings four specific upgrades over previous releases:
- Screen/UI understanding: Strong comprehension of digital screens across different devices.
- Grounding: Improved object detection and grounding via natural language queries.
- Multi-image input: Better reasoning when processing multiple images.
- Function calling: Stronger performance in both text-only and vision-text scenarios.
Training methodology
LFM2.5-VL-3B pairs a SigLIP2 400M NaFlex vision encoder with the same pre-trained backbone as the LFM2.5-2.6B text model. Pre-training used approximately 34T tokens, drawing 4x more vision data than before from curated and synthetic sets covering image captions, OCR, grounding, and instruction following. To support non-Latin scripts, the tokenizer was extended in place to double the vocabulary to 128K without retraining from scratch.
Post-training occurred in two stages. First was supervised fine-tuning (SFT), involving knowledge distillation from a larger teacher and Antidoom training. Second was multi-reward reinforcement learning (RL).
Performance results
Testing covered both vision and text benchmarks. The vision suite included multilingual visual comprehension, instruction following, visual math and scientific reasoning, document understanding, object detection, multi-image understanding, and screen understanding. The model leads its size class on real-world image tasks while reading digital content well, from documents and charts to on-screen UI elements.
| Task | Benchmark | LFM2.5-VL-3B (3.1B) | LFM2-VL-3B (3.1B) | gemma-4-E2B-it (5.1B) | gemma-4-E4B-it (8B) | InternVL 3.5 2B (2.4B) | InternVL 3.5 4B (4.7B) | Qwen3.5-2B (2.3B) | Qwen3.5-4B (4.7B) |
|---|---|---|---|---|---|---|---|---|---|
| General | MMStar | 63.3 | 57.7 | 45.3 | 52.9 | 57.7 | 65.5 | 55.1 | 59.3 |
| MME | 73.1 | 73.0 | 54.9 | 67.6 | 73.6 | 81.0 | 76.2 | 79.5 | |
| RealWorldQA | 73.1 | 71.1 | 60.0 | 64.3 | 61.6 | 67.7 | 65.1 | 67.1 | |
| SimpleVQA | 35.4 | 33.0 | 27.3 | 30.4 | 30.5 | 33.7 | 35.2 | 40.7 | |
| SEED-Bench (image) | 77.7 | 76.6 | 71.4 | 75.3 | 75.4 | 76.4 | 75.8 | 76.1 | |
| MMBench (dev EN v1.1) | 81.0 | 80.0 | 64.2 | 71.6 | 76.2 | 81.1 | 73.1 | 78.4 | |
| CountBenchQA | 87.3 | 92.2 | 70.4 | 80.5 | 70.4 | 82.5 | 83.8 | 86.7 | |
| Multilingual | MMMB | 83.0 | 81.9 | 73.3 | 80.4 | 76.3 | 81.5 | 75.9 | 82.0 |
| Multilingual MMBench | 79.5 | 76.3 | 62.8 | 71.2 | 70.9 | 76.6 | 69.9 | 77.0 | |
| Multimodal IF | MM-IFEval | 60.6 | 51.4 | 65.6 | 68.2 | 47.1 | 54.5 | 55.4 | 63.1 |
| STEM | LogicVista | 37.4 | 32.2 | 29.5 | 34.5 | 30.9 | 36.2 | 34.0 | 37.6 |
| MathVista (mini) | 68.5 | 62.1 | 37.8 | 45.2 | 56.8 | 67.1 | 48.7 | 63.6 | |
| MMMU-Pro | 30.5 | 28.7 | 26.9 | 32.6 | 21.3 | 22.7 | 24.9 | 36.0 | |
| MMMU (val) | 48.4 | 45.6 | 41.1 | 49.3 | 52.0 | 60.7 | 44.1 | 50.3 | |
| Document, OCR & Chart | ChartQA (test) | 81.3 | 80.4 | 43.2 | 42.1 | 81.7 | 86.2 | 78.4 | 84.2 |
| DocVQA (val) | 91.1 | 89.8 | 85.7 | 87.4 | 88.4 | 91.8 | 92.6 | 94.8 | |
| InfographicVQA (val) | 70.2 | 67.8 | 54.4 | 60.9 | 69.3 | 76.9 | 73.5 | 80.3 | |
| OCRBench v1 | 84.2 | 81.7 | 70.2 | 73.5 | 83.9 | 82.0 | 84.4 | 85.6 | |
| OCRBench v2 (En) | 47.5 | 43.9 | 44.4 | 48.8 | 45.5 | 49.1 | 47.7 | 58.7 | |
| TextVQA (val) | 84.3 | 83.0 | 62.5 | 69.0 | 76.6 | 77.5 | 77.3 | 81.2 | |
| Grounding | RefCOCO-avg | 87.9 | 57.1 | 67.3 | 72.1 | 82.9 | 88.8 | 78.5 | 86.6 |
| Multi-Image | BLINK | 61.5 | 50 Source Read original → Related articles![Do VLMs in production still use fixed-patch ViTs for their vision capabilities? [D]](https://ai-maestro.online/wp-content/uploads/2026/05/do-vlms-in-production-still-use-fixed-patch-vits-for-their-v-768x768.jpg)   |