LFM2.5-VL-3B: How to Use Its Vision Capabilities

Yesterday we released LFM2.5-VL-3B
and with it, a 𝗴𝘂𝗶𝗱𝗲 𝗼𝗻 𝗵𝗼𝘄 𝘁𝗼 𝘂𝘀𝗲 𝗶𝘁𝘀 𝘃𝗶𝘀𝗶𝗼𝗻 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝗶𝗲𝘀:
In this guide, you'll learn how to do:
• Single-image prompt
• Multi-image prompt
• OCR
• Document layout annotations
• Object detection and grounding
• Tool calling

Good news — this checks out. Here's the fact check:

Release: confirmed. Liquid AI released LFM2.5-VL-3B on August 12, 2026 (a 3.1B-parameter open-weight vision-language model), so "yesterday" checks out relative to a post on August 13–14.

The guide's six sections: confirmed exactly. Liquid AI's docs page "Vision Capabilities" (docs.liquid.ai/lfm/key-concepts/vision-capabilities) covers precisely these six topics, in this order:

1. Single-image prompt — captioning/VQA with one image
2. Multi-image prompt — labeled Media-1/Media-2 cross-references, flagged as best supported by the 3B model
3. OCR — reading text from document/screenshot images
4. Document layout annotations — structured region output (image_index=<n> <label> [xmin,ymin,xmax,ymax]) with labels like text/title/table/chart
5. Object detection and grounding — normalized [0,1000] bounding-box coordinates returned as JSON
6. Tool calling — image-conditioned function selection via chat-template tools=

Supporting facts on the model itself, also verified:

• Built on the LFM2.5-2.6B text backbone + SigLIP2 400M NaFlex vision encoder, ~34T pretraining tokens, 128K vocab
• RefCOCO precision@1 rose from 57.1 to 87.9, a 30-point gain over the previous release
• multi-image benchmarks improved (BLINK 50.2→61.5, MUIRBench 34.9→58.3)
• Available on Hugging Face (LiquidAI/LFM2.5-VL-3B) with native, GGUF, ONNX, and MLX formats, day-one support in llama.cpp/MLX/vLLM/SGLang/ONNX

One nuance worth flagging: the guide's individual sections note that multi-image, OCR-with-layout, grounding, and tool calling are "best supported by LFM2.5-VL-3B" specifically (as opposed to the smaller 450M/1.6B siblings in the same family) — so the framing that this is the guide for "its vision capabilities" is accurate, but it's really the flagship-tier guide within a three-model lineup. Otherwise, everything in your post lines up with the primary sources.