LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge

Today, we release LFM2.5-VL-3B, our most capable vision-language model. It delivers competitive vision performance against models twice its size, while running faster across a range of CPU and GPU deployments even compared with models that have fewer parameters.

LFM2.5-VL-3B builds on our previous LFM2-VL-3B, with significant improvements in screen understanding, grounding, function calling, and multi-image input. LFM2.5-VL-3B is a non-reasoning model that answers directly, keeping latency low for real-time and on-device applications.

The model is available today on Hugging Face and our Playground. Check out our docs on how to run and fine-tune it locally.

What’s New

LFM2.5-VL-3B extends the vision-language capabilities of our previous release with four major improvements:

Screen/UI understanding. LFM2.5-VL-3B has a strong understanding of digital screens across mobile, web, and desktop. It averages 80.7 on ScreenSpot-v2, far ahead of the much larger Gemma-4-E4B (51.2) and Qwen 3.5 4B (78.5) and close behind the larger InternVL-3.5-4B (84.1) 

Function calling. New to our VL line, LFM2.5-VL-3B is significantly stronger at tool use and function calling, for both text-only and vision-text inputs. ToolSandbox more than doubles from 26.4 to 59.5 and BFCL v4 climbs from 20.5 to 32.5, putting it on par with Gemma-4-E2B and ahead of Qwen3.5-2B.

Grounding. LFM2.5-VL-3B significantly improves grounding performance by scaling synthetic grounding data. It raises RefCOCO precision@1 from 57.1 to 87.9, a 30-point gain over our previous release.

Multi-image input. Improved reasoning across multiple images, improving BLINK from 50.2 to 61.5 and MUIRBench from 34.9 to 58.3.

Training

LFM2.5-VL-3B follows the architecture of our previous VLMs, LFM2.5-VL-1.6B and LFM2.5-VL-450M. It builds on the same pre-trained base as our just-released LFM2.5-2.6B text model, and integrates a SigLIP2 400M NaFlex encoder.

Compared to the preceding LFM2-VL-3B release, it is pre-trained on ~34T tokens. To better support non-Latin scripts in LFM2.5, we doubled the vocabulary to 128K by extending the existing tokenizer in place. 

Additionally, we scaled our vision pretraining by 4x tokens, while improving its data quality with a large mixture of curated and synthetic image-caption, OCR, grounding, and instruction following data.

The post-training pipeline consists of Supervised Fine-tuning (SFT), including knowledge distillation from a larger teacher model and Antidoom training, followed by multi-reward reinforcement learning (RL).

Vision Benchmarks

We evaluated LFM2.5-VL-3B across 28 benchmarks covering multilingual visual comprehension, instruction following, visual math and scientific reasoning, document understanding, object detection, multi-image understanding, and screen understanding.

LFM2.5-VL-3B is particularly strong at understanding both the physical and digital worlds. It leads its size class on real-world image tasks such as RealWorldQA, POPE, and RefCOCO, while also performing well on digital content, from documents and charts to on-screen UI elements. On average across these benchmarks, LFM2.5-VL-3B significantly outperforms much larger Gemma models and comes within 0.7% of the larger 4.7B Qwen 3.5 model.¹

Text Benchmarks

We also evaluated LFM2.5-VL-3B on text-only benchmarks for instruction following and tool use against the same set of models. Instruction following climbs across the board, and tool use improves sharply, with ToolSandbox more than doubling from 26.4 to 59.5. On tool use, LFM2.5-VL-3B is on par with Gemma-4-E2B and Qwen3.5-2B.²

These results make LFM2.5-VL-3B a strong, well-rounded vision-language model. It performs well across everyday tasks like image captioning, visual question answering, and document understanding, and it is especially strong at taking action, bounding box detection, and reading digital screens and documents. 

Fast inference everywhere

LFM2.5-VL-3B ships with day-one support across the inference ecosystem:

• llama.cpp — GGUF checkpoints for efficient edge inference
• MLX — Optimized inference for Apple Silicon
• vLLM — GPU-accelerated serving for production throughput
• SGLang — GPU-accelerated serving for production throughput
• ONNX — Cross-platform inference across diverse accelerators

On-device inference. LFM2.5-VL-3B decodes 228 tokens/s on an Apple M5 Max and 116 tokens/s on an AMD Ryzen AI Max+ 395, while staying in about 3 GB of memory. Its compact vision encoder also keeps it quick to first token, where models with heavier encoders such as InternVL stall on the same image. It even runs on a phone, decoding 20 tokens/s on a Galaxy S26 Ultra, so a capable vision model runs privately on your own device.

GPU inference. LFM2.5-VL-3B is built for fast, responsive inference on GPU. Because it answers directly instead of reasoning, it returns the first token quickly, even on image-heavy inputs. We measure time to first token on a single NVIDIA H100 SXM5 GPU, one request at a time, across three input types: a single 512x512 image, a 512x512 image with a 1,024-token input text, and 5 256x256 5-frame video clip. LFM2.5-VL-3B keeps latency consistently low and pulls far ahead on multi-frame inputs, returning the first token in about 34 ms on a 5-frame video clip where the Gemma models take around 200 ms.

We also measure output throughput (total output tokens / wall time) on a single NVIDIA H100 SXM5 GPU using a sustained-load setting: at each concurrency level, we continuously maintain the target number of in-flight requests, replacing each completed request immediately.

We benchmark each model with vLLM 0.26, using a 512×512 image plus 1,024 input tokens and up to 256 output tokens, in BF16, and report the median of 5 runs per concurrency level. LFM2.5-VL-3B reaches the highest output throughput of any model we tested, 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.

See LFM2.5-VL-3B in Action

You can try LFM2.5-VL-3B directly in your browser with our WebGPU demo. The model runs entirely on-device with no setup required. You can upload an image or capture a still image from your webcam to see the model in action.

Document understanding. Given a document page, the model analyzes the entire page in a single pass, identifying regions and assigning layout labels.

Screen/UI understanding. Given a high-level goal of finding a specific documentation page, the model navigates the website to satisfy the goal. The model visually analyzes each screen to identify where to click, and repeats this process until it reaches the target page.

Function calling. Given an image of a dish and a request for its recipe, the model identifies the dish and determines that it should call a search tool to find the recipe.

Grounding. Given a natural-language request, the model locates the specified objects in an image and returns a bounding box for each one.

Get Started

Start building today with LFM2.5-VL-3B, available on Hugging Face.

With LFM2.5, we're delivering on our vision of AI that runs anywhere. These models are:

• Open-weight — Download, fine-tune, and deploy without restrictions
• Fast from day one — Native support for llama.cpp, MLX, and vLLM across Apple, AMD, Qualcomm, and Nvidia hardware
• A complete family — From base models for customization to specialized audio and vision variants, one architecture covers diverse use cases

We can't wait to see what you build.

Download on Hugging Face

Read our docs

Try on Playground

Citation

Please cite this article as:

@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},
}

¹ 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.

All values in the table are normalized to 0–100. Original benchmarks use 0-2800 for MME, 0-1000 for OCRBench, and 0–1 for DocVQA and InfographicVQA.

RefCOCO-avg is average of P@1 over these 8 splits RefCOCO: val, testA, testB, RefCOCO+: val, testA, testB, RefCOCOg: val, test

POPE is F1 over all examples in adversarial, popular and random subsets

MMMU-Pro uses the 10-choice multiple-choice variant

² InternVL 3.5 models do not support function-calling.