LFM2.5-VL-3B release notes / what's new

Liquid AI has announced LFM2.5-VL-3B, a vision-language model (VLM) designed for high-performance on-device and edge deployment. The model is optimized for real-time applications, prioritizing direct answers over chain-of-thought reasoning to maintain low latency.

Key Capabilities and Improvements

LFM2.5-VL-3B introduces four primary enhancements over previous releases:

  • Screen and UI Understanding: Enhanced ability to interpret digital screens across various devices.
  • Grounding: Improved object detection and grounding using natural language queries.
  • Multi-image Input: Better reasoning capabilities when processing multiple images simultaneously.
  • Function Calling: Significant improvements in tool use and function calling for both text-only and vision-text scenarios.

Technical Architecture and Training

LFM2.5-VL-3B is a 3.1B parameter model that pairs a SigLIP2 400M NaFlex vision encoder with the pre-trained backbone used in the LFM2.5-2.6B text model.

Training Data and Process

  • Pre-training: The model was pre-trained on approximately 34 trillion tokens. This included four times more vision data than previous versions, utilizing curated and synthetic sets for image-captioning, OCR, grounding, and instruction-following.
  • Tokenizer: To support non-Latin scripts, the vocabulary was expanded to 128K using in-place tokenizer extension.
  • Post-training: The training pipeline consisted of two stages: supervised fine-tuning (SFT) utilizing knowledge distillation from a larger teacher model and Antidoom training, followed by multi-reward reinforcement learning (RL).

Performance Benchmarks

Vision Performance

LFM2.5-VL-3B leads its size class on real-world image tasks and demonstrates strong performance in reading digital content, including charts and UI elements. In comparative testing (normalized 0-100), it achieved a 69.4 average score across vision benchmarks, competing closely with InternVL 3.5 4B (69.4) and Qwen3.5-4B (70.1).

Notable benchmark results include:

  • ScreenSpot-v2 (Desktop/Mobile/Web): Significant jumps in performance over LFM2-VL-3B, with scores of 78.7 (Desktop), 81.2 (Mobile), and 82.2 (Web).
  • RefCOCO-avg (Grounding): Scored 87.9, a substantial increase from 57.1 in the previous version.
  • BLINK (Multi-Image): Scored 61.5, outperforming several larger models like Gemma-4-E4B-it (52.2) and Qwen3.5-2B (48.6).

Text and Tool Use Performance

On text-only benchmarks, LFM2.5-VL-3B showed improved instruction following (IFEval: 82.3) and a sharp increase in tool use capabilities. In ToolSandbox (59.5), it performed on par with Gemma-4-E2B (56.5) and Qwen3.5-2B (47.7).

Inference Speed and Hardware Compatibility

LFM2.5-VL-3B is designed for the edge, fitting into approximately 3 GB of memory. It provides day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX.

On-Device Benchmarks

  • M5 Max: 228 tokens/s
  • Ryzen AI Max+ 395: 116 tokens/s
  • Galaxy S26 Ultra: 20 tokens/s

GPU Throughput

On H100 GPUs, the model reaches output throughput of approximately 11K tokens per second at high concurrency, which is roughly double the throughput of 4B-class models. This allows for nearly 1 billion output tokens per day on a single H100.

Implementation

LFM2.5-VL-3B is available on Hugging Face and is compatible with transformers>=5.10.1. It can be loaded using AutoModelForImageTextToText and AutoProcessor for tasks involving image description, grounding, and OCR.

Sources

Related