Tasks and inputs
LFM3-VL-2B is listed for Image-text-to-text, Video-text-to-text, Document OCR. It accepts text, image, video and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
LFM3-VL-2B is Liquid AI's open 2B vision-language model for images and video on-device, strong on documents, charts and GUI screenshots.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under LFM Open License v1.0; no hosted price confirmed.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Screens are the target. Of all the categories Liquid AI measured for LFM3-VL-2B, GUI understanding (83.1) and document, OCR and chart reading (83.9) are where this 1.9B-parameter model shines, which makes it a natural fit for on-device assistants that read what is on a phone or laptop display. It was released on 27 October 2026 as the vision member of the LFM3 family, and it is the first Liquid vision model to accept video.
It is also the backbone of Liquid's d1-2B decision model.
| Component | Size |
|---|---|
| Language model (LFM3) | 1.73B |
| Vision encoder | 164M |
| Projector | 25M |
| Total | 1.92B |
The vision encoder is new and in-house: 14 layers mixing gated depthwise 2D convolutions with bidirectional attention, on 16x16 patches. Liquid says it uses about a third of the compute of SigLIP2. Images run at native resolution, each 2x2 patch block becoming one language token. Video is sampled at 2 fps, from 4 to 768 frames by default. Context is 128K tokens.
The model reasons in a <think> block before answering. With thinking, Liquid reports a 71.4 average across 62 vision benchmarks; without, 62.1. That nine-point gap is large, so leave thinking on unless latency forbids it. The average puts it close to Gemma 4 12B's 71.7 in Liquid's comparison.
Weak spots, by the same table: maths and logical reasoning (68.8), video (61.2) and agentic tasks (42.5). Liquid itself says it is not recommended for coding, multi-turn conversation or complex reasoning.
Recommended sampling: temperature 0.1, top_p 1.0, top_k 50, min_p 0, repetition penalty 1.1. Tool calling uses an XML format that is a superset of qwen3_coder, with deferred tool loading supported.
On a single B300 GPU it passes 12,000 tokens per second at 1,024 concurrent requests, so it is not only an edge model. Licence: LFM Open License v1.0.
Collect a hundred screenshots from your own app or workflow: forms, tables, dashboards, error dialogs. Write one factual question per image with a checkable answer ("what is the invoice total?"). Run with thinking on and off, measuring accuracy and time to answer on the target device. Add ten short screen recordings if video matters to you, since video is where it is weakest.
For a heavier alternative, compare Gemma 4 E4B IT. More under image-text-to-text and on Liquid's provider page.
Compare the model on the work you need to complete.
LFM3-VL-2B is listed for Image-text-to-text, Video-text-to-text, Document OCR. It accepts text, image, video and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under LFM Open License v1.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
sensetime
SenseNova-Loop-9B is SenseTime's 9B looped multimodal model: 32 layers run as 56, 512K context, ~15 GiB at 256K in 4-bit, open weights under Apache 2.0.
Tencent
Tencent Hy4 Flash is the low-cost Hy4 tier with open weights: a 220B MoE (~11.3B active) for agents and documents, with 1M context and image input.
shanghai-ai-lab
Sage S2.1 Flash is Shanghai AI Lab's open scientific multimodal agent model, reading molecules, crystals and proteins alongside text and images.
Meituan
LongCat-2.5 is Meituan's open-weight MIT-licensed multimodal MoE for long-horizon agents: screenshots, charts, video and multi-step coding.
sensetime
SenseNova-Loop-9B-Preview is SenseTime's looped 9B text-and-image model for long-context agentic work, running shared layers twice for extra depth.
NVIDIA
Nemotron 3.5 Super VL 120B A12B is NVIDIA's open multimodal reasoning MoE: 12B active, 1M context, image and video input, switchable thinking.