LFM3-Audio-2B

About

LFM3-Audio-2B is a compact, fully featured audio model that runs locally on laptops and edge devices: it takes text or audio as input and responds with text, speech, or interleaved text and speech. With 2B total parameters it is maintained and trained by LiquidAI (Liquid AI) and combines the 24-layer LFM3 backbone (13 gated short convolution blocks, 5 sliding-window attention, 6 global attention, grouped-query attention (GQA), partial rotary embeddings, sandwich norm, hidden size 2048) with two new audio components: a fully causal audio encoder - an LFM-based encoder mixing sliding-window attention and short convolution layers with bounded memory that supports live audio input and incremental prefill - and the in-house LiquidCodec audio tokenizer, a codebook model built on the LFM3 architecture that represents each 80 ms of speech as 12 discrete tokens (12 codebooks of size 2017, maskgit-style parallel decode with 3 MaskGIT steps in config) so generated tokens can be turned into speech at about 87x real time on a single CPU thread on an Apple M4 Max, with streaming decoding while speech is still being generated.

Capabilities. Voice chat: ask questions aloud and receive streaming replies in text and speech; for text-only answers the model can reason before answering via a <think> prefill (live voice chat is English only). Multilingual speech: transcribe, synthesize and translate across 18 speech languages / 19 locales (Arabic, Danish, German, Greek, British English, American English, Finnish, French, Hebrew, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Swedish, Turkish, Chinese). Audio understanding and reasoning: interpret recordings and answer questions about the speech, sounds and events they contain. Tool calling: turn spoken requests into Qwen3-style XML tool calls to control apps and devices, with text or speech alongside the calls. Special modes include LiveASR (model.transcribe_live(), streaming transcription in ~0.96 s chunks), TTS with female/male voices, and parallel speech prediction, where multiple audio tokens are predicted iteratively in parallel instead of sequentially - the detokenizer turns tokens into speech faster than real time for responsive conversations. Streaming performance: time to first token as low as 50 ms thanks to causal encoding and incremental prefill. Context length is 32K tokens (input + output); the audio encoder operates at 16 kHz while speech output is 24 kHz.

Performance (Liquid AI measurements): it improves over LFM2.5-Audio-1.5B in speech quality, spoken instruction following and audio understanding while adding multilingual speech and tool calling - English ASR WER 6.23 on the ESB / Open ASR suite, 6.03 multilingual ASR error on FLEURS (WER, CER for Chinese/Japanese/Korean), English TTS UTMOSv2 speech quality 3.83 (1-5 scale), VoiceBench 73.13, voice tool calling 58.50 (internal eval) and MMAR audio understanding 50.60.

Usage and deployment. Audio sampling parameters: audio_temperature=1.0, audio_top_k=128, audio_repetition_penalty=1.1, audio_repetition_window=4096 for speech synthesis and interleaved output; suppress_mixed=True and suppress_audio_out=True force text-only answers. Checkpoints ship as BF16 (vLLM, SGLang, Transformers), GGUF (llama.cpp), MLX 4-bit (MLX-VLM) and ONNX, with SFT fine-tuning notebooks (audio-to-function-calling, TTS) via Transformers. It is designed for local agents, voice assistants, ASR/TTS/translation, audio understanding, tool calling, structured outputs and domain-specific fine-tuning; not recommended for agentic coding or complex multi-step reasoning - use retrieval or search for knowledge-heavy tasks. Released October 27, 2026 (LFM3 release blog post) under the LFM Open License v1.0.