Is LFM3-VL-2B state of the art?Pending benchmarks.Announced Oct 27, 2026; no independent results yet.
Liquid AI's LFM3 vision-language model for on-device use. It takes text, images and video (the first LFM vision model with video input), can reason step by step before answering, and localizes objects for spatial and embodied reasoning. Open weights under the LFM Open License v1.0.
The details so far
- Lab
- Liquid AI
- Announced
- Oct 27, 2026
- API
- Live · Liquid AI API and Playground, per Liquid's blog
- Open weights
- Published
- Parameters
- 1.92B
- Architecture
- LFM3 language model (24 layers) plus a new in-house vision encoder (14 layers: 8 gated depthwise 2D convolutions, 6 bidirectional attention)
- Inputs
- Text, images and video at native resolution; video sampled at 2 fps
- Context
- 128K tokens (input + output)
- Languages
- Same 16 text languages as LFM3-1.7B
- License
- LFM Open License v1.0
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: across 62 vision benchmarks it matches Gemma 4 12B, surpasses Gemini 3.5 Flash Lite, is within 2 points of GPT-5.6 Luna, and leads similar-size models by at least 7.4 points.
Lab claim: the new vision encoder uses 3x less compute than the SigLIP2 encoder in earlier LFM VL models.
Sources
LABLiquid AI: Introducing LFM3liquid.ai/blog/lfm3 · Oct 27, 2026REPOLiquidAI/LFM3-VL-2B on Hugging Facehuggingface.co/LiquidAI/LFM3-VL-2BTracking continues until an independent board lists it.