Key Takeaways
- LFM2.5-VL-3B hits 20 tokens/s on a Galaxy S26 Ultra, making offline screen reading and tool use practical.
- It outscores the 5.1B Gemma-4-E2B by 18 points on ChartQA and nearly 50 on ScreenSpot-v2 Desktop.
- Open-weight release with quantized GGUFs and broad runtime support enables sovereign edge deployment without vendor lock-in.
Table of Contents
A 3B Vision Model That Runs on a Galaxy Phone — And Beats Gemma 4B
On August 12, LiquidAI published a 3.1‑billion‑parameter vision‑language model called LFM2.5‑VL‑3B that decodes 228 tokens per second on an M5 Max and fits into just 3 GB of memory.
A detailed post on the Hugging Face Blog reveals a small model that not only handles documents, charts, and UI screens at real‑time speed, but also outscores several larger competitors on hard vision benchmarks.
It achieves 20 tokens/s on a Galaxy S26 Ultra — a threshold that makes offline screen reading, object grounding, and tool calling genuinely practical without any cloud connection.
The jump from its predecessor is dramatic: screen‑understanding scores on ScreenSpot‑v2 went from 6.0 to 78.7 for desktop tasks, a 13‑fold improvement that repositions what edge AI can do with visual interfaces.
Inside LFM2.5-VL-3B: Screen Grounding, Multi-Image Reasoning, and Tool Use
The model pairs a SigLIP2‑based NaFlex vision encoder with the same LFM2.5‑2.6B text backbone that LiquidAI shipped earlier.
Pretraining consumed about 34 trillion tokens, with a fourfold increase in vision‑specific data spanning curated image‑caption pairs, OCR corpora, grounding sets, and synthetic instruction‑following examples.
To support non‑Latin scripts without rebuilding from scratch, the tokenizer was extended in place from 64K to 128K vocabulary entries — a clean engineering move that preserves compatibility while boosting multilingual coverage.
Post‑training runs in two stages: supervised fine‑tuning with knowledge distillation from a larger teacher and Antidoom training, followed by multi‑reward reinforcement learning.
Four capability pillars stand out.
Screen and UI understanding: the model reads digital interfaces across desktop, mobile, and web contexts with a coherence that eclipses every other model in its weight class.
Grounding: natural‑language object referencing hit 87.9 on RefCOCO average, up from 57.1 in the previous release, making it usable for real‑world agent workflows.
Multi‑image reasoning: scores on BLINK (61.5) and MuirBench (58.3) show it can cross‑reference visual information from several frames, a critical skill for document comparison or surveillance review.
Function calling: on ToolSandbox it jumps from 26.4 to 59.5, putting it on par with Gemma‑4‑E2B and Qwen3.5‑2B despite having fewer parameters, and enabling vision‑triggered API actions directly on‑device.
As the LiquidAI blog post details, benchmark tables confirm the edge: against the 5.1B Gemma‑4‑E2B, LFM2.5‑VL‑3B leads by 18 points on ChartQA, 16 on DocVQA, and nearly 50 points on ScreenSpot‑v2 Desktop.
Inference ecosystem support arrives on day one with llama.cpp, MLX, vLLM, SGLang, and ONNX.
Throughput on a single H100 reaches 11,000 tokens per second — roughly twice what 4B‑class alternatives deliver — translating to nearly 1 billion output tokens per day at high concurrency.
Edge AI Gains Ground: Where LiquidAI’s 3B VLM Fits in the Market
The official GGUF repository on Hugging Face already lists quantized versions as small as 1.48 GB (Q4_0), with the Q8_0 size at 2.73 GB and full‑precision BF16/F16 at 5.14 GB.
Over 5,300 downloads in the past month signal strong developer interest — and yet no cloud inference provider hosts the model, reinforcing that LiquidAI is betting entirely on local execution.
This stands in contrast to models like Apple FastVLM, Moondream 2, and SmolVLM2, which populate the small‑VLM landscape but often arrive with ecosystem constraints or restricted weight access.
LiquidAI’s open‑weight release, paired with aggressive quantization and broad runtime support, gives teams a sovereign edge‑deployment path that sidesteps vendor lock‑in.
The performance‑per‑watt story is undeniable: decoding 116 tokens/s on a Ryzen AI Max+ 395 means real‑time screen agents and document parsers can run on fanless hardware.
Yet the absence of a hosted API may slow early prototyping, a gap that dedicated fine‑tuning tutorials and a browser‑based WebGPU demo partially close.
As the edge VLM segment matures, models that combine speed, small footprint, and truly open distribution will pressure larger, cloud‑tethered alternatives to demonstrate why their extra parameters justify the latency.
What Developers Can Do Today With Offline Vision Intelligence
Downloadable GGUF weights and an instant WebGPU demo mean any developer can put a production‑grade vision model to work offline right now.
For teams building on‑device vision pipelines that need to scale, programmatic AI automation is how Andres SEO Expert moves from prototype to production — get in touch.
Frequently Asked Questions
What is LiquidAI LFM2.5-VL-3B and why is it significant?
LFM2.5-VL-3B is a 3.1-billion-parameter vision-language model by LiquidAI that runs efficiently on edge devices. It decodes 228 tokens per second on an M5 Max, fits in just 3 GB of memory, and outperforms larger competitors on several vision benchmarks, making offline screen reading, object grounding, and tool calling practical without a cloud connection.
How fast does LFM2.5-VL-3B run on different hardware?
According to the article, LFM2.5-VL-3B achieves 228 tokens per second on an M5 Max, 20 tokens per second on a Galaxy S26 Ultra, 116 tokens per second on a Ryzen AI Max+ 395, and up to 11,000 tokens per second on a single H100 GPU with high concurrency.
How does LFM2.5-VL-3B compare to Gemma 4B and other small vision models?
LFM2.5-VL-3B leads the 5.1B Gemma-4-E2B by 18 points on ChartQA, 16 on DocVQA, and nearly 50 points on ScreenSpot-v2 Desktop, despite having fewer parameters. It also outperforms other small models in screen understanding and grounding, while offering open weights and broad runtime support.
What can developers build with an offline 3B vision-language model?
Developers can build real-time screen agents, document parsers, object grounding systems, and vision-triggered API actions that run entirely on-device without cloud connectivity. The model’s speed and small footprint make it suitable for fanless hardware, mobile devices, and privacy-sensitive applications.
What benchmarks does LFM2.5-VL-3B excel at?
LFM2.5-VL-3B achieves strong scores across multiple benchmarks: 87.9 on RefCOCO average for grounding, 61.5 on BLINK and 58.3 on MuirBench for multi-image reasoning, 59.5 on ToolSandbox for function calling, and 78.7 on ScreenSpot-v2 Desktop for screen understanding.
How can developers access and deploy LFM2.5-VL-3B?
The model is available as open weights with GGUF quantizations as small as 1.48 GB (Q4_0). It supports llama.cpp, MLX, vLLM, SGLang, and ONNX on day one, and a WebGPU demo is available for instant browser-based testing. No cloud API is hosted yet, but the open distribution enables sovereign edge deployment.
What improvements does LFM2.5-VL-3B bring compared to its predecessor?
ScreenSpot-v2 Desktop scores jumped from 6.0 to 78.7, RefCOCO from 57.1 to 87.9, and ToolSandbox from 26.4 to 59.5. This was achieved by a fourfold increase in vision-specific pretraining data, an extended tokenizer for multilingual support, and a two-stage post-training pipeline with distillation and multi-reward reinforcement learning.
