# Liquid AI released LFM2-Audio as one compact audio-and-text foundation model Liquid AI released LFM2-Audio-1.5B on October 1, 2025, combining audio and text inputs and outputs in a 1.5B-parameter model. Canonical: https://rangoon.ai/insights/archive-liquid-lfm2-audio-october-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-10-01 Tags: [Liquid AI](https://rangoon.ai/insights/tags/liquid-ai/), [LFM2-Audio](https://rangoon.ai/insights/tags/lfm2-audio/), [speech AI](https://rangoon.ai/insights/tags/speech-ai/), [multimodal AI](https://rangoon.ai/insights/tags/multimodal-ai/) Source note: Historical reporting: Liquid AI released LFM2-Audio-1.5B on October 1, 2025 as an end-to-end audio-and-text model. The release states it can combine audio and text inputs and outputs and points to inference code plus a speech-to-speech reference implementation. A later LFM2 technical-report announcement describes speech processing as part of the family. Product capability and low-latency language are vendor claims, not independent production evidence. Image: A speech turn passes through a named audio-text model into a separately checked output mode. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - LFM2-Audio made input-output modality combinations a single-model design question. - The release’s reference implementation is a starting point, not a deployment receipt. - Speech, transcription, and audio generation still need separate task and failure tests. ## Liquid AI released LFM2-Audio-1.5B on October 1, 2025 [source 1](https://www.liquid.ai/blog/lfm2-audio-an-end-to-end-audio-foundation-model) Liquid AI announced LFM2-Audio-1.5B on October 1, 2025 as an end-to-end foundation model for audio and text generation. The launch calls the model 1.5B parameters and says it can switch among audio and text inputs and outputs. That combination is the specific release fact: it describes one backbone intended to cover conversational chat, transcription, text-to-speech, and audio classification. It does not mean every input-output pairing performs identically or is suitable for the same risk level. A speech transcript, spoken response, and audio classification label should each have their own quality and review criteria. ## One audio-text model still requires explicit input and output contracts [source 1](https://www.liquid.ai/blog/lfm2-audio-an-end-to-end-audio-foundation-model) The release frames LFM2-Audio as supporting all input-output combinations of audio and text. This is a model-level capability statement, not a complete application specification. A system must still decide sample rate, channel handling, speech segmentation, language selection, prompt placement, output format, and whether generated audio is played, stored, or reviewed. Those choices can alter both latency and meaning. A transcription task may need word timing and speaker labels; a speech-to-speech interaction may need interruption policy; a classifier may need calibrated confidence. The vendor description supports the family’s modality scope. It does not prove a team configured those contracts or measured their failure modes. ## Published inference code and a speech-to-speech example are distinct from a general deployment claim [source 1](https://www.liquid.ai/blog/lfm2-audio-an-end-to-end-audio-foundation-model) [source 2](https://www.liquid.ai/blog/lfm2-advancing-open-science-in-ai) Liquid says it published a Python package with inference code and a reference implementation for a speech-to-speech live-chat application. This is meaningful release material because it gives developers a stated starting path beyond a model name. It is not an assertion that every framework, device, or conversation flow is supported. Reference code can have assumptions about codec, buffering, model precision, hardware acceleration, and turn-taking that a different product must inspect. The later technical-report announcement identifies speech processing as part of the broader LFM2 work, but it is later context rather than launch-day proof of every implementation detail. A deployment should pin the code and model version and retain audio fixtures that cover silence, overlap, noise, and language changes. ## Low-latency language does not replace measured end-to-end speech behavior [source 1](https://www.liquid.ai/blog/lfm2-audio-an-end-to-end-audio-foundation-model) [source 2](https://www.liquid.ai/blog/lfm2-advancing-open-science-in-ai) Liquid presents LFM2-Audio as lightweight, multimodal, and real-time oriented. These are vendor positioning terms that describe the release objective. End-to-end delay, however, includes audio capture, packet or file handling, model preparation, inference, generation, playback, and any safety or logging step. A model can generate quickly after it receives clean audio while a product still feels slow or misses a speaker change. Testing should separately measure time to first partial output, final transcription, first generated audio, and recovery after an interruption. It should also compare the native audio route with a derived transcript route when both are offered. That evidence distinguishes a model capability announcement from a claim that a particular assistant has met a conversational service requirement. ## Sources 1. [Liquid AI: LFM2-Audio release](https://www.liquid.ai/blog/lfm2-audio-an-end-to-end-audio-foundation-model) 2. [Liquid AI: LFM2 technical report announcement](https://www.liquid.ai/blog/lfm2-advancing-open-science-in-ai) ## Related reading - [Liquid AI released LFM2-VL as compact vision-language models for edge deployment](https://rangoon.ai/insights/archive-liquid-lfm2-vl-edge-august-2025/) - [NVIDIA Riva 2.18 added Canary NIM options for offline multilingual speech work](https://rangoon.ai/insights/archive-nvidia-riva-canary-nim-february-2025/) - [NVIDIA released Parakeet-TDT-0.6B-v3 for 25-language speech-to-text](https://rangoon.ai/insights/archive-nvidia-parakeet-tdt-v3-august-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)