← All models

LFM3-Audio-2B

Liquid AI · 2026-10-27 · 2.1B parameters

LFM3-Audio-2B is Liquid AI's post-trained audio model in the first LFM3 family, released with open weights on October 27, 2026. It takes text or 16 kHz audio and answers in text, 24 kHz speech or interleaved text and speech, covering speech recognition, synthesis and translation in 18 languages, audio understanding and spoken tool calls, and it can reason before text-only answers. Its main weights hold 2,002,332,928 BF16 parameters: the LFM3 language model, a causal audio encoder of sliding-window attention and short convolutions, and heads that predict audio tokens in parallel; the repository also ships the 73,559,090-parameter LiquidCodec tokenizer, which represents each 80 ms of speech with 12 tokens, for 2,075,892,018 in the released checkpoint. The card states a 32K-token context while the backbone configuration allows 128,000 positions, so no exact context is recorded; the tokenizer assigns 125,033 distinct text IDs.

Benchmark scores