LFM3-Audio-2B-GGUF Audio-Text Multimodal Model: 4GB+ VRAM, File List

LFM3-Audio-2B-GGUF Audio-Text Multimodal Model: 4GB+ VRAM, File List

At a Glance

Item Value
Repository LiquidAI/LFM3-Audio-2B-GGUF
Family guide LFM3-Audio guide (1 articles)
Publisher guide Liquid AI: models and licenses
Published 2026-10-28
License lfm1.0
Formats GGUF
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

LiquidAI has released “LFM3-Audio-2B-GGUF", the GGUF version of their compact audio generation model “LFM3-Audio-2B". This model is designed for local execution on CPUs and GPUs in laptops and edge devices. It is an any-to-any model that allows free input of text or audio, and can generate responses in text, audio (speech), or an interleaved format of both.

Specifications

Below are the main specifications listed in the documentation for the original model “LFM3-Audio-2B" and its GGUF version.

  • Parameter count: Total parameters of 2B (2 billion parameters)
  • Architecture: LFM3 backbone, causal audio encoder, LiquidCodec audio tokenizer
  • Input specifications: Text or audio (audio encoder operates at 16 kHz)
  • Output specifications: Text, 24 kHz audio, or interleaved (alternating) output of text and audio
  • Context length: 32K tokens (combined limit for input and output)
  • Streaming capabilities: Supports causal audio encoding, incremental prefill, and streaming audio generation
  • Supported languages and locales: Arabic, Danish, German, Greek, British English, American English, Finnish, French, Hebrew, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Swedish, Turkish, Chinese
  • Recommended audio sampling settings: audio_temperature=1.0, audio_top_k=128, audio_repetition_penalty=1.1, audio_repetition_window=4096
  • License terms: LFM Open License v1.0 (lfm1.0)

Performance and Quality

Official benchmark comparison results are published in the model card for the source model “LFM3-Audio-2B". In the evaluation metrics, lower numbers for WER (Word Error Rate for transcription) and CER (Character Error Rate for transcription) indicate better performance.

→ Scroll horizontally to see all columns

Benchmark LFM3-Audio 2B LFM2.5-Audio 1.5B Gemma 4 E2B Gemma 4 E4B Qwen2.5-Omni 3B NVIDIA Audex 2B Voxtral Mini 3B
Total parameters 2B 1.5B 5B 8B 6B 3B 5B
English ASR WER ↓ 6.23 6.13 8.26 7.82 6.30 5.99 5.51
Multilingual ASR WER / CER ↓ 6.03 — 8.08 6.54 7.41 6.40 7.76
English TTS UTMOSv2 ↑ 3.83 3.22 — — 3.42 — —
VoiceBench ↑ 73.13 55.17 63.28 72.90 66.37 22.55 65.78
Tool calling internal eval ↑ 58.50 — 55.60 72.40 44.30 — 44.50
Audio understanding MMAR ↑ 50.60 25.85 44.49 49.10 54.41 53.21 50.10

According to the published figures, the original LFM3-Audio 2B model outperforms competing models and the previous-generation LFM2.5-Audio 1.5B in categories such as Multilingual ASR WER / CER (6.03), VoiceBench (73.13), and English TTS UTMOSv2 (3.83). On the other hand, its English ASR WER (6.23) falls short of Voxtral Mini 3B (5.51) and NVIDIA Audex 2B (5.99), and it also scores lower in Tool calling internal eval (58.50) and Audio understanding MMAR (50.60) compared to larger parameter models like Gemma 4 E4B (72.40) and Qwen2.5-Omni 3B (54.41).

Additionally, according to the performance description in the model card, the causal encoder design enables real-time processing of microphone input, with the time to first token reported to reach as fast as 50 ms. Furthermore, the audio decoder LiquidCodec is reported to operate at approximately 87x real-time speed on a single CPU thread of an Apple M4 Max, quickly converting generated tokens into audio.

Strengths and Use Cases

LFM3-Audio-2B is a versatile model designed to handle diverse audio tasks on edge devices and local environments. According to the original model’s documentation, it excels at the following specific use cases and features:

  • Real-time audio chat: When asked questions via voice, it can generate responses combining both text and audio in a streaming format. When answering with text only, it can also interleave reasoning processes before the response.
  • Multilingual audio processing and translation: It supports 18 languages and locales (Japanese, American English, British English, Chinese, Korean, German, French, Spanish, Italian, Dutch, Polish, Portuguese, Swedish, Turkish, Arabic, Danish, Greek, Finnish, Hebrew), allowing for Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and cross-lingual translation.
  • Live Automatic Speech Recognition (LiveASR): Transcribes speech sequentially while it is being spoken, without waiting for audio recording to finish. Currently limited to English support, it processes audio in real-time chunks of approximately 0.96 seconds.
  • Audio understanding and environmental sound interpretation: It analyzes not only speech content in recordings but also background noises, sound effects, and various event sounds, answering questions about their content.
  • Spoken tool calling: Generates JSON-formatted function calls from voice or text instructions, useful for smart home operations and application control.
  • Fine-tuning and domain specialization: Designed to easily perform additional training tailored to specific domains or tasks using Hugging Face Transformers.

On the other hand, due to its compact model size, it is not suited for complex multi-step reasoning or advanced coding tasks (agentic coding). For knowledge-intensive tasks, combining it with external search or Retrieval-Augmented Generation (RAG) is recommended.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 2.0B parameters (taken from the base model LiquidAI/LFM3-Audio-2B)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) BF16 3.2GB 3.9GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp as-is.

It is distributed in GGUF, so no conversion is needed.

License — lfm1.0 (Commercial use allowed with conditions): LiquidAI’s LFM Open License v1.0. Commercial use is allowed, except by legal entities with annual revenue of $10 million or more (Section 5). Redistribution requires the license text and notice of changes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Distributed Files

Weight files published in LiquidAI/LFM3-Audio-2B-GGUF, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
LFM3-Audio-2B-BF16.gguf 3.47GB
LFM3-Audio-2B-F16.gguf 3.47GB
LFM3-Audio-2B-Q4_0.gguf 1.05GB
LFM3-Audio-2B-Q4_K_M.gguf 1.10GB
LFM3-Audio-2B-Q8_0.gguf 1.85GB
mmproj-LFM3-Audio-2B-BF16.gguf 571MB
mmproj-LFM3-Audio-2B-F16.gguf 571MB
mmproj-LFM3-Audio-2B-Q8_0.gguf 338MB

How to Get It

GGUF-format weight files are available on the Hugging Face repository LiquidAI/LFM3-Audio-2B-GGUF. The model does not require individual consent (gated) and can be downloaded directly by anyone.

llama.cpp is supported as a runtime environment. However, according to the GGUF model card, official llama.cpp support for LFM3-Audio is awaiting the merge of a pull request (ggml-org/llama.cpp#31257), so it must be built and used from that pull request until it is merged and released.

When performing Automatic Speech Recognition (ASR), specify Perform ASR in <language>. or Perform ASR. for auto-detection in the system prompt, and use the llama-mtmd-cli command as follows:

llama-mtmd-cli -hf LiquidAI/LFM3-Audio-2B-GGUF:Q4_K_M \
  --audio speech.wav -sys "Perform ASR in English." -p "<__media__>" --temp 0

When performing Text-to-Speech (TTS), specify the language code (default is en) and voice type (female or male), and use the llama-tts command as follows:

llama-tts -m LFM3-Audio-2B-Q4_K_M.gguf -mm mmproj-LFM3-Audio-2B-Q8_0.gguf \
  -p "Hello world" --tts-voice male -o out.wav

Detailed option settings and usage instructions are also explained in the official llama.cpp deployment guide (https://docs.liquid.ai/deployment/on-device/llama-cpp).

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-10-27 LiquidAI MLX LiquidAI/LFM3-Audio-2B-MLX-4bit MLX 4bit 1.8GB (fits in 4GB VRAM)
2026-10-27 LiquidAI MLX LiquidAI/LFM3-Audio-2B-MLX-8bit MLX 8bit 2.6GB (fits in 4GB VRAM)
2026-10-27 LiquidAI MLX LiquidAI/LFM3-Audio-2B-MLX-bf16 MLX 4.4GB (fits in 8GB VRAM)

File sizes of each build:

  • Available builds in LiquidAI/LFM3-Audio-2B-MLX-4bit: MLX 4bit 1.5GB
  • Available builds in LiquidAI/LFM3-Audio-2B-MLX-8bit: MLX 8bit 2.2GB
  • Available builds in LiquidAI/LFM3-Audio-2B-MLX-bf16: MLX 3.7GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources