Liquid AI Unveils Six 2B-Scale LFM3 Models

Liquid AI has released six LFM3 models comprising text, vision, audio, omni, and decision models
All models are derived from LFM3-1.7B-Base, pre-trained on 22T tokens, and optimized for edge device execution with a size of approximately 1GB when 4-bit quantized
LFM3-1.7B scores an average of 49.37 across 29 benchmarks, higher than Qwen3.5-9B (47.24), and demonstrates strengths in local agents and tool calling
The KV cache size was directly calculated and verified, and it was announced that the model is faster and has a smaller KV cache than comparison models on the device

Model Composition and Performance

LFM3-VL-2B achieves an average score of 71.4 across 62 benchmarks, showing performance similar to Gemma 4 12B, and achieves a 2.2x reduction in compute using a proprietary 164M vision encoder
LFM3-Audio-2B supports voice processing in 18 languages and records a voice generation speed 87 times faster than real-time on the Apple M4 Max CPU
The decision model d1-2B scores 85.2 on Decision Bench, higher than the previous model d1-3B (83.6)

The application of LFM Open License v1.0 restricts commercial use by companies with annual revenue exceeding $10 million
Larger models are advantageous for coding, complex multi-step reasoning, and 128K context information retrieval, and RAG integration is recommended for knowledge work
Official support for major GPU serving frameworks such as SGLang and vLLM has not yet been merged, so caution is required when using them