LFM3-VL-2B
by LiquidAI
Liquid AI's 1.9B-parameter vision-language model for on-device use: text, images and video in, with step-by-step reasoning before it answers
LiquidAI/LFM3-VL-2BDeploy LFM3-VL-2B
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
LFM3-VL-2B is a vision-language model from Liquid AI, released in October 2026 for on-device visual understanding. It has 1.92B parameters (a 1.73B language model, a 164M vision encoder and a 25M projector) and takes text, images and video as input. It is the first LFM vision model to accept video, it reasons step by step before answering, and it locates objects and reasons about how they are placed. It is released under Liquid AI's LFM Open License 1.0.
Architecture
An LFM3 language model (24 layers: 13 gated short convolutions, 5 sliding-window attention and 6 global attention layers) built on LFM3-1.7B-Base, joined to a 164M-parameter vision encoder by a 25M-parameter projector.
Mixpeek SDK Integration
# Describe an image with LFM3-VL-2B (usage from the model card).
# pip install -U transformers torch
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "LiquidAI/LFM3-VL-2B"
model = AutoModelForImageTextToText.from_pretrained(
model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
messages = [{"role": "user", "content": [
{"type": "image", "url": "https://placecats.com/300/200"},
{"type": "text", "text": "Describe this image."},
]}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
return_dict=True, return_tensors="pt").to(model.device)
with torch.inference_mode():
output_ids = model.generate(**inputs, do_sample=True, temperature=0.1, top_k=50,
repetition_penalty=1.1, max_new_tokens=16384)
print(processor.batch_decode(output_ids[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])Capabilities
- Questions about images and recorded video
- Step-by-step reasoning on harder visual questions such as charts and math
- Object localization and spatial reasoning
- 16 languages listed on the model card
- Runs with Hugging Face Transformers or behind an OpenAI-compatible server
Use Cases on Mixpeek
Performance
We have not measured it. Its reasoning can be long, which is why the model card's example sets max_new_tokens to 16,384; budget generation time accordingly on device.
Common Pipeline Companions
Frequently Asked Questions
What is LFM3-VL-2B?
A 1.92B-parameter vision-language model from Liquid AI, built for on-device use. It takes text, images and video, reasons step by step before it answers, and can locate objects and reason about where they sit. It was released in October 2026 under the LFM Open License 1.0.
Can LFM3-VL-2B understand video?
Yes. It is the first model in Liquid AI's LFM vision line to take video as input, so you can ask about a recorded lecture, a walkthrough filmed on a phone or a clip from a camera feed.
Can I use LFM3-VL-2B commercially?
It is published under Liquid AI's LFM Open License 1.0, listed as a custom licence on Hugging Face. Read its terms for your company size and use before shipping it in a product.
How do I use it with Mixpeek?
Run it where your files are to caption images or video segments, then upsert the captions with the file and timestamp into a collection so they are searchable. To search the visual content directly instead, use a multimodal embedding model in Mixpeek's managed extractor.
Specification
Build a pipeline with LFM3-VL-2B
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free