Join waitlist
Back to products
LFM2-Audio

LFM2-Audio

Real-time audio conversations on-device

Website liquid.ai
Overview

What it is

LFM2 by Liquid AI is a new class of open foundation models designed for on-device speed and efficiency. Its hybrid architecture delivers 2x faster CPU performance than Qwen3 and SOTA results in a tiny footprint.

Intent

I need it when

Reduce infrastructure costs and training time for custom audio foundation models

LFM2-Audio achieves 3x faster training efficiency compared to previous generation, with flexible deployment across CPU, GPU, and NPU hardware, lowering the cost-per-capability for audio AI systems

Deploy efficient AI models on edge devices for real-time audio processing with minimal latency

LFM2-Audio enables on-device audio inference with 2x faster decode and prefill performance on CPU compared to competing models, delivering millisecond latency and data sovereignty for audio workloads on phones, IoT devices, and embedded systems

Build private, offline-capable audio AI applications without sending data to cloud servers

LFM2-Audio runs locally on-device through LEAP platform and Liquid Apollo, ensuring audio data never leaves the device while maintaining competitive performance across 0.35B, 0.7B, and 1.2B parameter sizes

Access open-source audio models for research, prototyping, and non-commercial deployment

LFM2-Audio weights are freely available on Hugging Face, llama.cpp, and OpenRouter with no commercial restrictions for organizations under $10M revenue, enabling rapid experimentation and community-driven audio AI development

Drop

Not a fit when

  • User requires hosted API inference without self-hosting or custom deployment infrastructure
  • Organization generates over $10M annual revenue and needs commercial licensing without custom negotiation
  • Use case demands real-time cloud-based audio processing with guaranteed uptime SLAs
  • Project requires proprietary audio codec support or specialized audio format handling beyond standard inference
  • User needs pre-built audio-specific features like speech recognition, voice cloning, or audio generation without custom integration
Commercials

Pricing

Free open-source with optional enterprise and startup support tiers View pricing