On October 27, Liquid AI released the LFM3 family, including six models spanning language, vision, speech, and structured decision-making. Built on a 1.7-billion-parameter foundation with specialized versions around 2 billion parameters, the models can be customized on a single GPU and deployed on devices, private servers, or cloud. The text model API is priced at $0.08 per million input tokens and $0.20 per million output tokens.
Liquid AI announced the launch of d1-ultra, a decision model it claims to be the world's best, saying it outperforms Jev 1.14 in eight task categories and surpasses competitors such as vLLM Semantic Router d3 overall.
Liquid AI released LFM2.5-24B-A2B, the largest model in the LFM2.5 series, an efficient local model for laptops and PCs with 24B total parameters and 2B active per token. It features a 131K-token context, a 128K-token vocabulary, 33T pretraining tokens, reasoning capabilities, and easy fine-tuning. It performs competitively with Qwen3.6-35B-A3B on agentic and tool-use tasks, and its base and post-trained weights are open-sourced on Hugging Face.
Liquid AI's D1 decision model is now live on OpenRouter. It accepts application state and answers yes/no, choice, or score questions, returning structured answers with probabilities for each option. It supports zero data retention, with pricing of $0.04 per million input tokens, free output, and a 65K context length.
Liquid AI released an experimental DSpark draft model for its vision-language model LFM2.5-VL-3B. It uses speculative decoding to speed up inference without changing output quality, achieving up to 3.13x decode speedup on device and 2.66x on H100, with end-to-end gains up to 2.62x and 2.27x. The drafter adds 280M parameters, an 8.9% increase, and supports llama.cpp, MLX-VLM, and SGLang from day one.