# Liquid AI released LFM2-VL as compact vision-language models for edge deployment Liquid AI released LFM2-VL on August 12, 2025, introducing 450M and 1.6B vision-language models with conditional open-license terms. Canonical: https://rangoon.ai/insights/archive-liquid-lfm2-vl-edge-august-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-08-12 Tags: [Liquid AI](https://rangoon.ai/insights/tags/liquid-ai/), [LFM2-VL](https://rangoon.ai/insights/tags/lfm2-vl/), [multimodal AI](https://rangoon.ai/insights/tags/multimodal-ai/), [edge AI](https://rangoon.ai/insights/tags/edge-ai/) Source note: Historical reporting: Liquid AI released LFM2-VL on August 12, 2025 as its first vision-language family, with 450M and 1.6B variants. The release says the models accept text and image inputs and are available through Hugging Face with example fine-tuning code. Liquid labels its license open and based on Apache 2.0 while describing commercial-use conditions. Vendor speed and benchmark figures are not independent validation. Image: A selected vision model receives an image and text prompt under a defined device budget. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - LFM2-VL added text-and-image input to Liquid’s small edge-model line. - The 450M and 1.6B variants are separate artifacts with different capability and resource tradeoffs. - An Apache-based license with revenue conditions needs term-by-term review. ## Liquid AI released its first LFM2 vision-language family on August 12, 2025 [source 1](https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models) Liquid AI released LFM2-VL on August 12, 2025 as a vision-language family for low-latency, device-aware deployment. The release names a 450M model for tighter resource settings and a 1.6B model for a larger but still lightweight option. It says the family accepts text and image inputs at variable resolutions. This establishes a dated multimodal model event and a stated input boundary. It does not establish that an arbitrary camera feed, PDF page, or high-resolution image will fit a chosen device budget. Image decoding, tiling, pixel limits, prompt construction, output length, and runtime precision still determine what a local deployment can do. ## A 450M and 1.6B release is a model-selection decision, not one uniform capability [source 1](https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models) The launch separates LFM2-VL-450M from LFM2-VL-1.6B and publishes benchmark rows for both. That makes the release more concrete than a generic vision feature: a team can choose between named checkpoints. It also makes broad family claims less useful. A smaller model can change image-detail retention, OCR behavior, visual grounding, and generation reliability, while a larger model can change load time and memory demand. Liquid’s results are vendor measurements under its stated setup. They should not be read as a guarantee for a device, image source, or user population. A comparison needs fixed images, resolutions, token budgets, model revisions, and acceptance criteria for the task at hand. ## The release describes an open license with a commercial-use threshold [source 1](https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models) Liquid says LFM2-VL is under an open license based on Apache 2.0 and allows academic and research use, plus commercial use for companies below a stated revenue threshold. It directs larger companies to contact the vendor for a commercial license. This wording is a release-specific legal boundary, not a detail to erase in an “open model” label. A user must examine the actual license, the organization’s status, distribution plan, and any upstream component terms. Public weights do not prove unrestricted downstream redistribution, data rights, or hosted-service rights. The release identifies a path to obtain the models; it does not grant an unqualified opinion on a particular organization’s use. ## Liquid’s speed and benchmark tables need their workload conditions to travel with them [source 1](https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models) Liquid reports benchmark and timing comparisons for LFM2-VL against other compact vision models. It specifies a typical workload for its speed discussion: one 1024-by-1024 image, a short prompt, and 100 generated tokens. That is valuable method context because an image model’s apparent speed depends on visual encoding and text generation together. A deployment that uses multiple pages, longer prompts, constrained JSON, or high-resolution crops may have a different bottleneck. The event supports a claim that Liquid published a small vision-language family and vendor evaluations. It does not establish independent performance, safety, or accuracy for a product workflow. A local report should retain the selected resolution and output policy beside its own measurements. ## Sources 1. [Liquid AI: LFM2-VL release](https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models) 2. [Liquid AI: LFM2 technical report announcement](https://www.liquid.ai/blog/lfm2-advancing-open-science-in-ai) ## Related reading - [Liquid AI released LFM2-Audio as one compact audio-and-text foundation model](https://rangoon.ai/insights/archive-liquid-lfm2-audio-october-2025/) - [Liquid AI released LFM2 as a small on-device foundation-model family](https://rangoon.ai/insights/archive-liquid-lfm2-on-device-july-2025/) - [MedGemma 1.5 expanded a 4B model toward volumes, slides and longitudinal imaging inputs](https://rangoon.ai/insights/archive-google-medgemma-1-5-imaging-january-2026/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)