Liquid AI Releases LFM2.5-VL-3B, a 3B-Class Vision-Language Model Built for On-Device Deployment

Liquid AI Releases LFM2.5-VL-3B

Liquid AI has released LFM2.5-VL-3B, a multimodal vision-language model designed for on-device deployment. The model builds on the company's earlier LFM2-VL-3B checkpoint with what Liquid AI describes as additional mid- and post-training, targeting improved grounding, object detection, and document OCR.

Architecture and Specs

LFM2.5-VL-3B pairs the LFM2.5-2.6B language model backbone with a SigLIP2 NaFlex vision encoder (400M parameters), which processes images at native resolution by splitting large images into non-overlapping 512×512 patches alongside a resized whole-image thumbnail. The model has a vocabulary of 128,000 tokens and a 32,768-token context window. It supports 16 languages, including English, Arabic, Chinese, French, German, Japanese, Korean, and Hindi.

Pricing is not applicable — the model is distributed as open weights on Hugging Face under the LiquidAI organization, available in four formats: the original checkpoint (for Transformers, vLLM, and SGLang), a quantized GGUF export (for llama.cpp CPU inference), an ONNX export (for cross-platform hardware-accelerated deployment), and an MLX export optimized for Apple Silicon.

Performance Claims

According to Liquid AI, LFM2.5-VL-3B runs at 228 tokens per second on an Apple M5 Max and 116 tokens per second on an AMD Ryzen AI Max+ 395, using under 3.3 GB of memory. The company states the model delivers improved grounding and object detection through natural language queries, along with full-page OCR that outputs layout annotations — labeling regions as text, titles, tables, charts, equations, and more, with normalized bounding-box coordinates and OTSL (Optimized Table Structure Language) formatting for tables.

No standardized benchmark scores (e.g., MMLU, DocVQA) were disclosed in the release materials reviewed for this article.

Intended Use

Liquid AI recommends the model for single-turn, high-throughput, low-latency tasks — for example, near-real-time object detection in automotive applications, batch OCR processing of scanned documents to convert PDFs into searchable text, and on-device translation of signs and menus. The company explicitly advises against using it for long-context or reasoning-intensive tasks such as visual web design or technical blueprint analysis.

The model supports function calling via a Pythonic tool-call format and uses a ChatML-like chat template consistent with the broader LFM2.5 model family.

What This Means

LFM2.5-VL-3B continues Liquid AI's focus on efficient, edge-deployable models rather than competing directly with frontier cloud models on raw capability. The sub-3.3 GB memory footprint and stated throughput on consumer hardware (Apple M5 Max, AMD Ryzen AI Max+ 395) position this as a candidate for offline document processing, automotive vision systems, and mobile translation apps rather than general-purpose chat or agentic coding. The lack of published benchmark comparisons against similarly sized open models (e.g., Qwen2-VL-2B, InternVL2) makes it difficult to independently verify Liquid AI's grounding and OCR improvement claims — buyers evaluating the model for production use should run their own domain-specific tests before committing.