Liquid AI Releases LFM2.5-VL-3B Vision-Language Model for Edge and On-Device Deployment

Liquid AI Releases LFM2.5-VL-3B Vision-Language Model for Edge and On-Device Deployment Liquid AI has launched LFM2.5-VL-3B, an edge-focused vision-language model designed to run on consumer hardware while providing direct answers for real-time applications. The 3.1-billion-parameter model introduces enhanced screen and UI comprehension, grounding, multi-image reasoning, and function calling. Architecture couples a SigLIP2 400M NaFlex vision encoder with an LFM2.5-2.6B backbone, trained on approximately 34T tokens with an expanded 128K vocabulary. Post-training utilizes supervised fine-tuning with teacher distillation and Antidoom training, followed by multi-reward reinforcement learning. Benchmark evaluations show an average score of 69.4 across vision benchmarks, matching or exceeding several larger 4B–8B parameter models in document understanding, grounding, and GUI tasks. Hardware performance reaches 228 tokens/s on Apple M5 Max, 20 tokens/s on Galaxy S26 Ultra, and about 11,000 tokens/s output throughput on a single GPU under high concurrency. Architecture and Training Pipeline LFM2.5-VL-3B integrates a SigLIP2 400M NaFlex vision encoder with the pre-trained backbone of the LFM2.5-2.6B text model. Pre-training incorporated around 34T tokens with four times the vision data of previous releases, encompassing synthetic and curated OCR, image-caption, instruction-following, and grounding datasets. The tokenizer vocabulary was doubled to 128K via in-place expansion to support non-Latin scripts. Post-training was conducted in two sequential stages: supervised fine-tuning incorporating knowledge distillation from a larger teacher alongside Antidoom training, followed by multi-reward reinforcement learning. Multimodal and Text Performance Across multimodal evaluations, LFM2.5-VL-3B achieved an overall average score of 69.4, competitive with or exceeding 4B- to 8B-parameter baselines such as Gemma-4 and InternVL 3.5. Notable benchmark results include 87.9 on RefCOCO-avg grounding, 91.1 on DocVQA, and over 78 to 82 on ScreenSpot-v2 desktop, mobile, and web suites. In text evaluations, the model achieved 82.3 on IFEval and 59.5 on ToolSandbox, matching or surpassing similarly sized models in instruction following and tool use. Inference and Deployment Ecosystem Designed to fit inside approximately 3 GB of memory, LFM2.5-VL-3B supports on-device deployment across systems including llama.cpp, MLX, vLLM, SGLang, and ONNX. On-device benchmarks demonstrate decoding speeds of 228 tokens/s on Apple M5 Max, 116 tokens/s on AMD Ryzen AI Max+ 395, and 20 tokens/s on a Samsung Galaxy S26 Ultra. Under high concurrency on a single GPU, the model achieves around 11,000 output tokens per second, equivalent to nearly 1 billion output tokens daily on an H100 accelerator.

Liquid AI发布边缘端视觉语言模型LFM2.5-VL-3B Liquid AI推出专为边缘端优化的视觉语言模型LFM2.5-VL-3B。该模型采用直接回答而非逐步推理的模式以保障实时与端侧响应速度，在屏幕与UI解析、视觉定位、多图推理和工具调用等能力上较前代显著提升。 模型结合了SigLIP2 400M NaFlex视觉编码器与LFM2.5-2.6B文本骨干，预训练规模约34T tokens，词表扩充至128K以支持非拉丁字符。 在综合视觉基准测试中平均得分为69.4，并在屏幕理解（ScreenSpot-v2）和定位（RefCOCO-avg）等任务上较前代有大幅飞跃。 端侧解码速度在M5 Max上可达228 tokens/s，内存占用约3 GB，高并发GPU推理输出吞吐量达到约1.1万 tokens/s。 架构与两阶段后训练 LFM2.5-VL-3B将SigLIP2 400M NaFlex视觉编码器与LFM2.5-2.6B骨干网络结合，在约34T tokens上预训练，视觉数据量增加至过去的4倍，涵盖图像字幕、OCR、目标定位和指令微调数据集。研发团队通过就地扩展分词器将词表增至128K以支持非拉丁文本。后训练分为两个阶段：先采用监督微调（SFT），结合大模型知识蒸馏和Antidoom训练；随后进行多奖励强化学习（RL）。 多模态与纯文本基准测试表现 在各家3B级模型主流评测中，LFM2.5-VL-3B在真实图像任务和界面解析方面表现突出。其多图BLINK得分为61.5，RefCOCO-avg目标定位从前代的57.1跃升至87.9，桌面端ScreenSpot-v2评分从6.0飙升至78.7。在纯文本评测中，IFEval得分为82.3，ToolSandbox工具调用得分为59.5，与同规模竞品持平或领先。 端侧部署与推理吞吐表现 该模型自发布起即适配llama.cpp、MLX、vLLM、SGLang与ONNX等推理框架。硬件表现方面，其内存占用仅约3 GB，在Ryzen AI Max+ 395上解码为116 tokens/s，Galaxy S26 Ultra上达20 tokens/s。在GPU环境下，模型保持了较低的多帧延迟，并在高并发测试中达到约11K tokens/s的吞吐量，相当于单张H100每天可产出接近10亿输出tokens。

LFM2.5-VL-3B is our most capable vision-language model you can run on your own hardware. It understands documents and screens alike, grounds objects, and can call tools. It answers directly instead of reasoning, so responses stay fast in real-time and on-device apps. LFM2.5-VL-3B extends the vision-language capabilities of our previous releases with four major improvements: Screen/UI understanding: Strong understanding of digital screens across different devices. Grounding: Improved grounding and object detection with natural language queries. Multi-image input: Improved reasoning across multiple images. Function calling: Significantly stronger at function calling, in text-only and vision-text situations. How we trained our most capable vision-language model LFM2.5-VL-3B pairs a SigLIP2 400M NaFlex vision encoder with the same pre-trained backbone as our LFM2.5-2.6B text model. It is pre-trained on about 34T tokens, with 4x more vision data than before, drawn from curated and synthetic image-caption, OCR, grounding, and instruction-following sets. To support non-Latin scripts, we doubled the vocabulary to 128K by extending the tokenizer in place rather than retraining from scratch. Post-training runs in two stages: First is supervised fine-tuning (SFT), with knowledge distillation from a larger teacher and Antidoom training . Second is multi-reward reinforcement learning (RL). Benchmark results We evaluated LFM2.5-VL-3B across both vision and text benchmarks. The vision benchmarks cover multilingual visual comprehension, instruction following, visual math and scientific reasoning, document understanding, object detection, multi-image understanding, and screen understanding. LFM2.5-VL-3B leads its size class on real-world image tasks, while also 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 50.2 45.2 52.2 52.0 57.2 48.6 58.7 MuirBench 58.3 34.9 32.9 51.8 45.0 53.5 48.2 62.0 Hallucination HallusionBench 47.2 46.4 41.8 49.8 47.6 52.1 49.3 51.7 POPE 88.7 89.2 84.0 86.9 88.0 88.9 88.6 86.0 GUI ScreenSpot-v2 Desktop 78.7 6.0 28.1 45.8 79.9 82.0 63.8 76.3 ScreenSpot-v2 Mobile 81.2 7.6 42.9 60.3 86.2 87.8 69.7 81.4 ScreenSpot-v2 Web 82.2 2.5 22.4 47.6 79.9 82.6 65.9 77.8 Average - 69.4 57.2 52.0 59.7 64.6 69.4 63.7 70.1 *All values in the table are normalized to 0–100. Evaluation is done using vLLM 0.26.0 and each model’s recommended generation parameters when available. Non-reasoning mode is used everywhere, and models are prompted to directly answer without reasoning. We also evaluated LFM2.5-VL-3B on text-only benchmarks for instruction following and tool use. Instruction following climbs across the board, and tool use improves sharply. On tool use, LFM2.5-VL-3B is on par with Gemma-4-E2B and Qwen3.5-2B. 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) Instruction following IFEval 82.3 72.9 83.0 87.9 32.4 35.4 73.6 86.2 IFBench 25.8 20.8 34.1 39.2 24.4 24.5 28.9 33.5 Multi-IF 59.4 46.5 69.4 77.4 16.3 16.9 53.5 66.7 Tool use & function calling ToolSandbox 59.5 26.4 56.5 61.6 N/A N/A 47.7 65.0 BFCL V4 32.5 20.5 33.2 40.0 N/A N/A 33.9 53.6 *InternVL 3.5 models do not support function-calling. These results demonstrate that LFM2.5-VL-3B is a strong, general-purpose vision-language model. It covers everyday tasks (captioning, visual question answering, document understanding) and is especially good at grounding objects, reading screens and documents, and calling tools. Inference speed on CPU and GPU LFM2.5-VL-3B ships with day-one support across the inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX. On-device inference. LFM2.5-VL-3B decodes 228 tokens/s on an M5 Max and 116 tokens/s on a Ryzen AI Max+ 395, and fits in about 3 GB of memory. It even reaches 20 tokens/s on a Galaxy S26 Ultra, so you can run it fully on-device. GPU inference. LFM2.5-VL-3B keeps latency consistently low and is the fastest on multi-frame inputs. LFM2.5-VL-3B is also the fastest on output throughput out of all models we tested, reaching about 11K tokens per second at high concurrency. That is roughly 2× the larger 4B-class models and ahead of even the smaller 2B-class models, which adds up to nearly 1B output tokens per day on a single H100. How to use LFM2.5-VL-3B Reach for LFM2.5-VL-3B when you need on-device intelligence for high-volume workloads. Install the latest version of transformers (compatible with transformers>=5.0.0 ): % pip install -q torch torchvision accelerate "transformers>=5.10.1" Then load and run the model: import torch from transformers.image_utils import load_image from transformers import AutoModelForImageTextToText, AutoProcessor from IPython.display import display MODEL_ID = "LiquidAI/LFM2.5-VL-3B" processor = AutoProcessor.from_pretrained(MODEL_ID) model = AutoModelForImageTextToText.from_pretrained( MODEL_ID, device_map= "auto" , dtype= "bfloat16" , ) img_url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/coco_sample.png" input_image = load_image(img_url) display(input_image) messages = [ { "role" : "user" , "content" : [ { "type" : "image" , "image" : input_image}, { "type" : "text" , "text" : "Describe this image in two concise sentences." }, ], } ] inputs = processor.apply_chat_template( messages, add_generation_prompt= True , tokenize= True , return_dict= True , return_tensors= "pt" , ).to(model.device) with torch.inference_mode(): outputs = model.generate( **inputs, do_sample= True , temperature= 0.2 , top_k= 50 , repetition_penalty= 1.0 , max_new_tokens= 256 , ) output = processor.batch_decode(outputs[:, inputs[ "input_ids" ].shape[ 1 ]:], skip_special_tokens= True )[ 0 ] print (output) Two cats are sleeping on a pink couch with two remote controls. You can find more hands-on examples on how to use LFM2.5-VL3B for multi-image inputs, grounding, OCR, tool calling, and more in our documentation . Check out our release blog for video examples. LFM2.5-VL-3B demo Check out this browser demo of LFM2.5-VL-3B powering a vision-capable chat interface . It allows you to take or upload multiple images and let the model interact with them, including grounding, OCR, and tool use. Get Started LFM2.5-VL-3B is available on Hugging Face today. With LFM2.5, we're delivering on our vision of AI that runs anywhere. These models are: Download: LFM2.5-VL-3B on Hugging Face. Try: run the WebGPU demo in your browser , no setup needed. Fine-tune: adapt LFM2.5-VL-3B to your task with our fine-tuning tutorials . We can't wait to see what you build. Citation Please cite this article as: Liquid AI, "LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge", Liquid AI Blog, Aug 2026. Or use the BibTeX citation: @article{liquidAI2026VL3B, author = {Liquid AI}, title = {LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge}, journal = {Liquid AI Blog}, year = {2026}, note = {www.liquid.ai/blog/lfm2-5-vl-3b}, }