LFM3-VL-2B

Multimodal
Liquid AI

LFM3-VL-2B is Liquid AI's open-weight 1.92B vision-language model: a 1.73B LFM3 language backbone plus a new 164M in-house convolution/attention vision encoder and a 25M projector, released on October 27, 2026 as part of the LFM3 family. It accepts text, images and video (the first LFM model with video input, sampled at 2 fps) and answers in text, with a thinking mode used for the published scores. Trained on 22T tokens, published under the LFM Open License v1.0, with weights on Hugging Face for llama.cpp, MLX, vLLM and ONNX.

Key Specifications

Parameters
1.9T
Context
-
Release Date
October 27, 2026
Average Score
70.7%

Timeline

Key dates in the model's history
Announcement
October 27, 2026
Last Update
November 3, 2026
Today
November 20, 2026

Technical Specifications

Parameters
1.9T
Training Tokens
22.0T tokens
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Multimodal

Working with images and visual data
ChartQA
Vendor table (ChartQA (test)); thinking mode • Self-reported
92.7%
DocVQA
Vendor table (DocVQA (val)); thinking mode • Self-reported
96.7%

Other Tests

Specialized benchmarks
ActivityNet
Vendor table (ActivityNet-QA (test)); thinking mode • Self-reported
71.2%
BLINK
Vendor table (BLINK); thinking mode • Self-reported
71.1%
CharXiv-D
Vendor table (CharXiv (descriptive, val)); thinking mode • Self-reported
87.5%
CharXiv-R
Vendor table (CharXiv (reasoning, val)); thinking mode • Self-reported
70.2%
CountBench
Vendor table (CountBenchQA); thinking mode • Self-reported
93.0%
DynaMath
Vendor table (DynaMath (average accuracy)); thinking mode • Self-reported
79.2%
EmbSpatialBench
Vendor table (EmbSpatial-Bench); thinking mode • Self-reported
82.9%
ERQA
Vendor table (ERQA); thinking mode • Self-reported
48.2%
Hallusion Bench
Vendor table (HallusionBench); thinking mode • Self-reported
65.6%
HR-Bench (4k)
Vendor table (HR-Bench (4K)); thinking mode • Self-reported
82.3%
InfoVQA
Vendor table (InfographicVQA (val)); thinking mode • Self-reported
86.2%
LVBench
Vendor table (LVBench); thinking mode • Self-reported
41.2%
MathVision
Vendor table (MATH-Vision (full)); thinking mode • Self-reported
51.6%
MathVista-Mini
Vendor table (MathVista (mini)); thinking mode • Self-reported
84.2%
MLVU-M
Vendor table (MLVU (dev, M-Avg)); thinking mode • Self-reported
71.3%
MMBench-V1.1
Vendor table (MMBench (dev EN v1.1)); thinking mode • Self-reported
86.1%
MME
Vendor table (MME); thinking mode • Self-reported
86.9%
MM IF-Eval
Vendor table (MM-IFEval); thinking mode • Self-reported
72.8%
MMMU-Pro
Vendor table (MMMU-Pro); thinking mode • Self-reported
45.7%
MMMU (val)
Vendor table (MMMU (val)); thinking mode • Self-reported
55.7%
MMStar
Vendor table (MMStar); thinking mode • Self-reported
74.4%
MMVU
Vendor table (MMVU (val)); thinking mode • Self-reported
52.9%
MotionBench
Vendor table (MotionBench (dev)); thinking mode • Self-reported
64.0%
MuirBench
Vendor table (MuirBench); thinking mode • Self-reported
66.4%
MVBench
Vendor table (MVBench); thinking mode • Self-reported
69.4%
OCRBench
Vendor table (OCRBench v1); thinking mode • Self-reported
86.3%
OCRBench-V2 (en)
Vendor table (OCRBench v2 (En)); thinking mode • Self-reported
63.5%
POPE
Vendor table (POPE); thinking mode • Self-reported
88.7%
RealWorldQA
Vendor table (RealWorldQA); thinking mode • Self-reported
79.3%
RefCOCO-avg
Vendor table (RefCOCO-avg); thinking mode • Self-reported
90.5%
RefSpatialBench
Vendor table (RefSpatial-Bench); thinking mode • Self-reported
49.8%
RoboSpatialHome
Vendor table (RoboSpatial-Home); thinking mode • Self-reported
63.4%
ScreenSpot Pro
Vendor table (ScreenSpot-Pro); thinking mode • Self-reported
55.5%
SimpleVQA
Vendor table (SimpleVQA); thinking mode • Self-reported
36.3%
TextVQA
Vendor table (TextVQA (val)); thinking mode • Self-reported
87.9%
VideoMME w/o sub.
Vendor table (Video-MME (v1, no subtitles)); thinking mode • Self-reported
62.8%
VideoMME w sub.
Vendor table (Video-MME (v1, subtitles)); thinking mode • Self-reported
69.0%
VideoMMMU
Vendor table (Video-MMMU); thinking mode • Self-reported
49.6%
We-Math
Vendor table (We-Math (testmini, strict)); thinking mode • Self-reported
68.6%

License & Metadata

License
lfm1_0
Announcement Date
October 27, 2026
Last Updated
November 3, 2026

Compare LFM3-VL-2B

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.