Skip to content
Neural Daily

ModelsNews

Liquid AI LFM3: six open 2B models on one backbone

🇭🇺 Olvasd el magyarul →

Liquid AI released the LFM3 family: text, vision, audio, omni and decision models, all around 2B parameters and built on one shared base model. All are open-weight and fine-tune on a single GPU.

By: Mikola Máté Patrik · · 3 min read

LFM3 model family diagram showing how LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B and d1-Omni-2B branch from LFM3-1.7B-Base
Image: Liquid AI · Press kit · Official diagram from the Liquid AI release blog · Source

What was released?

Liquid AI released the LFM3 model family on Tuesday (Hugging Face). There are six models, all branching from the same LFM3-1.7B-Base, so you pick the model that fits your modality instead of running one large generalist. All weights are on Hugging Face, including the base model for your own post-training.

  • LFM3-1.7B: text, 256K context, tool calling, built for local agents.
  • LFM3-VL-2B: image and video input, the first Liquid vision model with video.
  • LFM3-Audio-2B: speech recognition, speech synthesis, translation across 18 languages, tool calls from speech.
  • LFM3-Omni-2B: image, video and audio in, text and speech out, the company's first omni model.
  • d1-2B and d1-omni-2B: decision models that answer several questions in one forward pass with calibrated probabilities.

What changed from LFM2?

The new architecture has 24 layers: 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. It adds head-wise gated attention, partial RoPE and sandwich norm. According to Liquid, it is about 25% faster than LFM2-1.2B with 8K tokens in context. The base model was trained on 22T tokens, with the final 3.5T tokens carrying a larger share of agentic traces and tool calls.

LFM3-1.7B-Base architecture diagram with gated short convolution and gated GQA blocks and the 24-layer pattern
Liquid AI · Press kit · Official architecture diagram from the Liquid AI release blog · Source

Source: Liquid AI. LFM3-1.7B-Base architecture: 13 gated short convolution, 6 global attention and 5 sliding-window attention layers.

Liquid's numbers

  • LFM3-1.7B: beats LFM2.5-2.6B with 35% fewer parameters, scoring 38.42 on WebArena-Verified against 3.02. About 1 GB at 4-bit.
  • LFM3-VL-2B: average across 62 vision benchmarks rises from 57.6 to 71.4 versus LFM2.5-VL-3B, on par with Gemma 4 12B according to Liquid.
  • New vision encoder: 3x less compute than the previous SigLIP2 encoder.
  • Audio: time to first token as low as 50 ms, and LiquidCodec generates speech at about 87x real time on one CPU thread of an Apple M4 Max.
  • LFM3-Omni-2B: beats Qwen2.5-Omni-7B on audio-visual video understanding, per Liquid.
  • d1-2B: replaces d1-3B with a third fewer parameters, scoring 75.5 against 64.0 on ImajevBench.

These are Liquid's own results and have not been independently verified yet.

What to keep in mind

The license is not Apache or MIT but Liquid's own LFM Open License 1.0, so read it before commercial use. The models run in transformers but require trust_remote_code, and the audio and omni variants need a recent transformers version.

Why it matters

If you build on-device agents, voice assistants or a robotics perception layer, LFM3 gives you a small base for every modality that you can fine-tune on one GPU. Liquid also published Colab notebooks for SFT, DPO, GRPO and CPT fine-tuning, so it is quick to try on your own data.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles