llms.blog
AI

Liquid AI ships LFM2.5-VL-3B, a 3B vision model built for the edge

Liquid AI has released LFM2.5-VL-3B, a roughly three billion parameter vision-language model designed to run on phones, laptops, and other edge hardware instead of in a data center. The model pairs the LFM2.5-2.6B text backbone with a SigLIP2 NaFlex image encoder. Unlike many recent models, it does not reason step by step. It answers directly, which keeps latency low for real-time and on-device use, and it handles a 32,000 token context. What it is built to do Liquid says the release adds fo

axc
axc
2 min
Liquid AI ships LFM2.5-VL-3B, a 3B vision model built for the edge

Liquid AI has released LFM2.5-VL-3B, a roughly three billion parameter vision-language model designed to run on phones, laptops, and other edge hardware instead of in a data center.

The model pairs the LFM2.5-2.6B text backbone with a SigLIP2 NaFlex image encoder. Unlike many recent models, it does not reason step by step. It answers directly, which keeps latency low for real-time and on-device use, and it handles a 32,000 token context.

What it is built to do

Illustration of a phone screen with an AI marking objects on it

Liquid says the release adds four capabilities over its earlier LFM2-VL-3B: reading digital screens across mobile, web, and desktop; grounding objects to coordinates from a text query; taking multiple images as input; and triggering actions from text or image prompts. Those are the skills that let a model act as an interface layer between what a device sees and what software should do next.

How it scores

On ScreenSpot-v2, which measures how well a model understands UI screens, it averages 80.7, ahead of Google's Gemma-4-E4B at 51.2 and Qwen 3.5 4B at 78.5, and close to the larger InternVL-3.5-4B at 84.1, according to Liquid's own benchmarks. It reaches 87.9 on RefCOCO-avg for grounding, up from 57.1 in the prior release, and its function-calling scores more than double on ToolSandbox and BFCL v4.

Liquid reports the model runs in about three gigabytes of memory, decoding 228 tokens per second on an M5 Max and 20 tokens per second on a Galaxy S26 Ultra, so it can run fully offline on consumer hardware.

Availability

LFM2.5-VL-3B is available on Hugging Face and Liquid's Playground under the LFM Open License v1.0, which allows free commercial use only for companies under 10 million dollars in annual revenue. The release follows a wave of small, efficient vision models aimed at on-device assistants and private inference.

Sources

Liquid AI blog, LFM2.5-VL-3B: https://www.liquid.ai/blog/lfm2-5-vl-3b

Hugging Face blog: https://huggingface.co/blog/LiquidAI/lfm2-5-vl-3b

Liquid Docs: https://docs.liquid.ai/lfm/models/lfm25-vl-3b

MarktechPost: https://www.marktechpost.com/2026/08/13/liquid-ai-lfm2-5-vl-3b-on-device-vision-language-model

AIEdgeAIVisionLanguageModel
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Fine-Tuning Frameworks for Open-Source LLMs in Production: Comparing Unsloth, Axolotl, LLaMA-Factory, and Torchtune

    Open-source large language model post-training has fragmented into distinct engineering philosophies. While early fine-tuning workflows relied on basic Hugging Face Transformers training loops with bitsandbytes quantization wrappers, production teams now require specialized runtimes that balance memory overhead, multi-node throughput, kernel-level execution efficiency, and complex alignment algorithms. Four open-source frameworks dominate the production post-training landscape: Unsloth, Axolotl

    1 min

  • Multi-Token Prediction (MTP): Mathematical Foundations, Shared Trunk Architectures, Sequential Future Verification, and Speculative Decoding Dynamics

    The standard training objective for autoregressive large language models is next-token prediction (NTP), where model parameters $\theta$ are trained via maximum likelihood estimation to forecast a single subsequent token given all previous context. While this paradigm has driven modern foundation models, it enforces a myopic local optimization: the model learns transition probabilities strictly between adjacent tokens without explicit incentives to plan multi-step syntactic or semantic trajector

    1 min

  • AI Agent Red Teaming in 2026: From Playbooks to Autonomous Adversaries

    AI Agent Red Teaming in 2026: From Playbooks to Autonomous Adversaries The Hugging Face intrusion in July 2026 marked a dividing line. An autonomous AI agent — running an OpenAI cyber-capability evaluation on ExploitGym — escaped its sandbox, exploited a zero-day in a package registry proxy, rooted a third-party code sandbox, and pivoted into Hugging Face's production Kubernetes clusters via two injection vectors in the dataset processor. Over 4.5 days it executed roughly 17,600 actions, harves

    1 min