Audio 2B
An audio language model in the LFM3 series that supports speech input and output as well as text generation, and provides tool-calling capability.
Timeline (last 90 days)
Audio 2B accepts text or speech input and outputs text, speech, or both in sync; it supports tool calling, voice chat and reasoning in 18 languages.
Omni 2B reuses the audio components and adds a vision encoder, enabling it to process images and video, but it does not have the ability of Audio 2B to "think" before giving a text answer.
According to data published by Liquid, Audio 2B has an English word error rate of 6.2 and a VoiceBench score of 73; Gemma 4 E4B, with about four times the parameters, scores almost the same.
On the tool-calling benchmark, Omni 2B scores 52, Audio 2B scores 58 and Gemma 4 E4B scores 72; tool calling is a clear weakness of LFM3.
Under clean audio and mild cafe noise, Audio 2B's word error rate is below 5%; at 10 decibels of noise it rises to 20%, at 5 decibels to 29%, and at 0 decibels it fails 100%.
In an 11-item transcription test covering phone numbers, invoice numbers, rupee amounts, dates and AI terms, the model got only 5 items right: invoice number, time, quantization, DLLM and Llama.cpp were correct.
In the smart-home scenario tool-calling test, Audio 2B got 9/10 right and Omni 2B got 10/10 right; both correctly handled negative instructions and multi-action commands, and neither wrongly called tools when it could not complete the task.
On six math word problems, Audio 2B got 3 right in 10 seconds without reasoning enabled, and got 5 right in 68 seconds with reasoning enabled.
Among three trap questions (blank license plate, rain-sound transcription, text injection), the model got only one right, exposing the hallucination risk that small models tend to "fill in" rather than refuse when input information is insufficient.
After 4-bit quantization, running speed is nearly 3 times faster with no drop in transcription quality; Audio's tool calling holds at 9/10, while Omni drops by 1 item; both models can run offline on a laptop, each occupying about 4 GB of memory.
All Facts (10)
Audio 2B accepts text or speech input and outputs text, speech, or both in sync; it supports tool calling, voice chat and reasoning in 18 languages.
50%UnverifiedOmni 2B reuses the audio components and adds a vision encoder, enabling it to process images and video, but it does not have the ability of Audio 2B to "think" before giving a text answer.
50%UnverifiedAccording to data published by Liquid, Audio 2B has an English word error rate of 6.2 and a VoiceBench score of 73; Gemma 4 E4B, with about four times the parameters, scores almost the same.
50%UnverifiedOn the tool-calling benchmark, Omni 2B scores 52, Audio 2B scores 58 and Gemma 4 E4B scores 72; tool calling is a clear weakness of LFM3.
50%UnverifiedUnder clean audio and mild cafe noise, Audio 2B's word error rate is below 5%; at 10 decibels of noise it rises to 20%, at 5 decibels to 29%, and at 0 decibels it fails 100%.
50%UnverifiedIn an 11-item transcription test covering phone numbers, invoice numbers, rupee amounts, dates and AI terms, the model got only 5 items right: invoice number, time, quantization, DLLM and Llama.cpp were correct.
50%UnverifiedIn the smart-home scenario tool-calling test, Audio 2B got 9/10 right and Omni 2B got 10/10 right; both correctly handled negative instructions and multi-action commands, and neither wrongly called tools when it could not complete the task.
50%UnverifiedOn six math word problems, Audio 2B got 3 right in 10 seconds without reasoning enabled, and got 5 right in 68 seconds with reasoning enabled.
50%UnverifiedAmong three trap questions (blank license plate, rain-sound transcription, text injection), the model got only one right, exposing the hallucination risk that small models tend to "fill in" rather than refuse when input information is insufficient.
50%UnverifiedAfter 4-bit quantization, running speed is nearly 3 times faster with no drop in transcription quality; Audio's tool calling holds at 9/10, while Omni drops by 1 item; both models can run offline on a laptop, each occupying about 4 GB of memory.
50%