---

license: apache-2.0

base_model:

• LiquidAI/LFM2.5-VL-3B

language:

• en

pipeline_tag: image-text-to-text

library_name: transformers

tags:

• text-generation-inference
• llama-cpp
• 3B

---

LFM2.5-VL-3B-GGUF

> LFM2.5-VL-3B is Liquid AI's multimodal, on-device-optimized vision-language model, built on the LFM2-VL-3B foundation with further mid- and post-training, combining the LFM2.5-2.6B language backbone with a SigLIP2 NaFlex shape-optimized 400M vision encoder that processes images at native resolution by splitting large images into non-overlapping 512×512 patches plus a resized thumbnail. It supports a 32,768-token context window across 16 languages, delivers improved grounding and object detection with natural-language queries, and introduces full-page OCR with structured layout annotation (bounding boxes, region labels like text/table/equation, LaTeX for formulas, and OTSL for tables). It's designed for single-turn, high-throughput, low-latency tasks — near-realtime object detection, batch OCR of scanned documents, or on-device translation of menus and signs — rather than long-context or reasoning-intensive work like visual web design. Compared to its LFM2-VL-3B predecessor, it shows substantial gains in screen understanding (80.7 on ScreenSpot-v2, versus far behind on tool-use benchmarks previously), grounding (87.9 on RefCOCO), multi-image reasoning (58.3 on MuirBench), and tool use (59.5 on ToolSandBox), competing closely with larger models like Gemma4-E4B while running at 228 tok/s on an Apple M5 Max and 116 tok/s on an AMD Ryzen AI Max+ 395 in under 3.3GB of memory (and even 20 tok/s on a Galaxy S26 Ultra), or reaching ~11K tokens/sec throughput on a single H100 via vLLM. It's available in native, GGUF, ONNX, and MLX formats, supports Pythonic tool calling, and is released under Liquid AI's LFM1.0 license.

Model Files

File Name | Quant Type | File Size | File Link |

|-----------|------------|-----------|-----------|

| LFM2.5-VL-3B.BF16.gguf | BF16 | 5.4 GB | Download |

| LFM2.5-VL-3B.F16.gguf | F16 | 5.4 GB | Download |

| LFM2.5-VL-3B.Q3_K_L.gguf | Q3_K_L | 1.45 GB | Download |

| LFM2.5-VL-3B.Q3_K_M.gguf | Q3_K_M | 1.37 GB | Download |

| LFM2.5-VL-3B.Q3_K_S.gguf | Q3_K_S | 1.27 GB | Download |

| LFM2.5-VL-3B.Q4_K_M.gguf | Q4_K_M | 1.67 GB | Download |

| LFM2.5-VL-3B.Q4_K_S.gguf | Q4_K_S | 1.6 GB | Download |

| LFM2.5-VL-3B.Q5_K_M.gguf | Q5_K_M | 1.94 GB | Download |

| LFM2.5-VL-3B.Q5_K_S.gguf | Q5_K_S | 1.9 GB | Download |

| LFM2.5-VL-3B.Q8_0.gguf | Q8_0 | 2.87 GB | Download |

| LFM2.5-VL-3B.mmproj-bf16.gguf | mmproj-bf16 | 856 MB | Download |

| LFM2.5-VL-3B.mmproj-f16.gguf | mmproj-f16 | 856 MB | Download |

| LFM2.5-VL-3B.mmproj-q8_0.gguf | mmproj-q8_0 | 583 MB | Download |

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp