LFM-VL

weights

open(model.safetensors on the Hub for every LFM2.5-VL size, ungated)

data

described(the release post describes a mixture of curated and synthetic caption, OCR, grounding and instruction data

code

partial(LoRA fine-tuning notebooks in the Liquid4All cookbook

license

LFM-Open-License-1.0(Apache-2.0-based

LFM2.5-VL's weights can be used commercially by companies with under 10 million dollars in annual revenue; larger companies need a paid license from Liquid AI. Liquid publishes fine-tuning notebooks but neither its training data nor its training code. LFM-Open-License-1.0 allows commercial use only within a bound, so the license tier is use_bounded.

• https://huggingface.co/api/models?author=LiquidAI&search=VL&limit=1000&expand[]=downloads&expand[]=createdAt&expand[]=lastModified&expand[]=cardData&expand[]=gated recorded 2026-09-27

  All 31 LiquidAI VL checkpoints carry license other with license_name lfm1.0, none gated

• https://huggingface.co/api/models/LiquidAI/LFM2.5-VL-1.6B?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified&expand[]=siblings recorded 2026-09-27

  Hub metadata for LFM2.5-VL-1.6B: gated false, model.safetensors and LICENSE, license_name lfm1.0

• https://huggingface.co/api/models/LiquidAI/LFM2.5-VL-3B?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified&expand[]=siblings recorded 2026-09-27

  Hub metadata for LFM2.5-VL-3B: gated false, model.safetensors and LICENSE, license other, license_name lfm1.0

• https://huggingface.co/LiquidAI/LFM2.5-VL-3B/raw/main/LICENSE recorded 2026-09-27

  LFM Open License v1.0: "\"Threshold\" shall mean annual revenue of 10 million United States dollars ($10,000,000) or more"; "Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed under this Agreement"

• https://huggingface.co/LiquidAI/LFM2.5-VL-3B/raw/main/README.md recorded 2026-09-27

  Fine-tuning table: "SFT (Unsloth)" and "SFT (TRL)", LoRA notebooks; no training code linked

• https://www.liquid.ai/blog/lfm2-5-vl-3b recorded 2026-09-27

  "Today, we release LFM2.5-VL-3B, our most capable vision-language model"; "a large mixture of curated and synthetic image-caption, OCR, grounding, and instruction following data"

• https://www.liquid.ai/lfm-license recorded 2026-09-27

  "Once you cross the revenue threshold, you must contact Liquid AI, Inc. to purchase a commercial license to continue using the model"

Hugging Face downloads over the trailing 30 days, summed across Liquid AI's LFM2-VL and LFM2.5-VL checkpoints, including its own GGUF, ONNX and MLX builds. More than a third of it is those quantized builds, which suit local apps.

• https://huggingface.co/api/models?author=LiquidAI&search=VL&limit=1000&expand[]=downloads&expand[]=createdAt&expand[]=lastModified&expand[]=cardData&expand[]=gated recorded 2026-09-27

  Thirty-one LiquidAI VL checkpoints with 308,517 downloads in the trailing 30 days, 116,217 of them GGUF, ONNX and MLX builds; LFM2-VL-1.6B 83,372

LFM2.5-VL reads full pages, charts and screens and answers over several images, and the 3B model reports strong chart and OCR results for its size. Liquid demonstrates video captioning but reports no video benchmark, so it sits level with Moondream rather than with MiniCPM-V.

• https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B/raw/main/README.md recorded 2026-09-27

  "improved results on multi-image inputs, high-resolution images, and OCR"; a real-time video stream captioning WebGPU demo; no video benchmark

• https://huggingface.co/LiquidAI/LFM2.5-VL-3B/raw/main/README.md recorded 2026-09-27

  "LFM2.5-VL-3B significantly improves over LFM2-VL-3B in screen understanding, grounding, multi-image input, and tool use"; "Full page OCR with layout annotation"; MMStar 63.3, ChartQA 81.3, MuirBench 58.3, ScreenSpot-v2 80.7

• https://www.liquid.ai/blog/lfm2-5-vl-3b recorded 2026-09-27

  OCRBench v1 84.2; latency quoted "on a 5-frame video clip"; no video benchmark reported