Tasks and inputs
LFM3-Audio-2B is listed for Audio-to-audio, Automatic speech recognition, Text-to-speech. It accepts text, audio and produces text, audio. Use the publisher’s input format and preprocessing for your chosen deployment.
LFM3-Audio-2B is Liquid AI's open 2B speech model: ASR, TTS, voice chat and interleaved text-and-speech output from one local checkpoint.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under LFM Open License v1.0; no hosted price confirmed.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Text, speech, or both at once. LFM3-Audio-2B can answer a spoken question with interleaved output, streaming words and 24 kHz audio from the same model, which removes the separate recogniser and synthesiser most voice apps chain together. Liquid AI released it on 27 October 2026 as the audio member of the LFM3 family, aimed at laptops and edge devices.
Its Hugging Face repo existed from September and went public with the launch.
The English-only restriction on live ASR and voice chat is the detail to catch early if you serve other markets.
A causal audio encoder (sliding-window attention plus short convolutions, 16 kHz input) sits on the LFM3 backbone. Output speech goes through LiquidCodec, Liquid's tokenizer, at 12 tokens per 80 ms with streaming decode. LiquidCodec decodes at about 87x real time on a single CPU thread of an Apple M4 Max. Time to first token can be as low as 50 ms with streaming input. Context is 32K tokens, input plus output.
| Benchmark | Score | Note |
|---|---|---|
| English ASR WER | 6.2 | Voxtral Mini 3B 2507 is better at 5.5 |
| Multilingual ASR WER/CER | 6.0 | Best in Liquid's table |
| English TTS UTMOSv2 | 3.8 | Best in Liquid's table |
| VoiceBench | 73.1 | Best in Liquid's table |
| Tool calling (internal) | 58.5 | Gemma 4 E4B leads at 72.4 |
Liquid's own figures. Notice that English ASR is essentially flat against LFM2.5-Audio-1.5B (6.1); the gains are in conversation and synthesis.
Liquid's launch post says 18 languages. The model's speech list has 19 entries because it counts British and American English separately. Same thing, different counting. Research elsewhere put it at 19; the 18-language figure is the cleaner one.
BF16 for vLLM, SGLang and Transformers (5.20.2+, with trust_remote_code=True, plus librosa and soundfile), GGUF for llama.cpp, MLX 4-bit through MLX-VLM, and ONNX. Liquid does not recommend it for agentic coding or complex multi-step reasoning.
Record thirty realistic spoken requests in a noisy room. Measure end-of-speech to first audio, check that interleaved answers stay coherent when text and audio diverge, and have a person rate the voice. Repeat the ASR portion in every language you need, since voice chat will not help outside English.
See also audio-to-audio models and Liquid's provider page.
Compare the model on the work you need to complete.
LFM3-Audio-2B is listed for Audio-to-audio, Automatic speech recognition, Text-to-speech. It accepts text, audio and produces text, audio. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 32.77K context. Compare pronunciation, voice consistency and time to the first playable audio. Check the service’s billing unit for the actual script.
The weights are downloadable under LFM Open License v1.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
rime
Rime Prism is a speech-to-speech model for enterprise voice agents that settles each reply as inspectable text first, with parallel tool calls.
Alibaba
Qwen-Audio-3.1-Realtime Flash is Alibaba's lighter full-duplex speech model with audio and text in and out, function calling, web search and a 262K context.
ElevenLabs
Eleven Realtime v1 Max is ElevenLabs' reasoning speech-to-speech agent for support and multi-step tool workflows, OpenAI Realtime-compatible.
ElevenLabs
Eleven Realtime v1 Mini is ElevenLabs' fast, low-cost speech-to-speech voice agent stack, with an OpenAI Realtime-compatible API and 90+ languages.
xAI
Grok Voice 3.0 is xAI's speech-to-speech model for real-time voice agents, replacing Grok Voice 2.0 at the same $0.08-per-minute price.
Amazon
Nova 2.5 Sonic is Amazon's speech-to-speech model on Bedrock for real-time voice agents, with async tool calling, polyglot voices and a 256K context.