Liquid AI releases LFM2-2.6B small model
Liquid AI has released LFM2-2.6B, the largest model yet in its LFM2 series, which the company says outperforms several 3B–4B open models while using a hybrid architecture built for faster inference.
Liquid AI announced LFM2-2.6B, the newest and largest model in its Liquid Foundation Model 2 (LFM2) family. It joins the existing 350M, 700M and 1.2B variants.
The company says the 2.6-billion-parameter model, trained on 10 trillion tokens, outperforms several larger models: Llama 3.2-3B-Instruct, Gemma-3-4b-it and SmolLM3-3B. Liquid reports 82.41% on GSM8K (math reasoning) and 79.56% on IFEval (instruction following). The announcement does not give competitor scores or other benchmark details, and the results have not been independently verified.
Liquid says it tuned the model specifically for English and Japanese. It claims strong performance in French, Spanish, German, Italian, Portuguese, Arabic, Chinese and Korean as well.
The model uses a hybrid architecture that alternates Grouped Query Attention (GQA) blocks with short convolutional layers. According to Liquid, this design gives faster inference and needs less KV cache than other modern architectures, which should lower deployment costs. The post does not include specific latency or memory figures.
LFM2-2.6B is available on Hugging Face under Liquid's own LFM Open License, not a standard open-source license. The post does not describe the license terms.
Liquid says it plans to keep scaling its foundation models to run on more devices. It did not give a timeline or name future model sizes.