Auguments/LiquidAI-LFM2.5-Audio-1.5B-GGUF-BF16 overview
LFM2.5 Audio 1.5B GGUF Highest quality quant original LiquidAI's GGUFs are FP16, which may degrade quality by clipping values. Or maybe not, but why risk it? I…
Runs locally from ~133.8 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| LFM2.5-Audio-1.5B-BF16.gguf | GGUF | BF16 | 2.18 GB | Download |
| LFM2.5-Audio-1.5B-Q6_K.gguf | GGUF | Q6_K | 918.2 MB | Download |
| mmproj-LFM2.5-Audio-1.5B-BF16.gguf | GGUF | BF16 | 437.6 MB | Download |
| tokenizer-LFM2.5-Audio-1.5B-BF16.gguf | GGUF | BF16 | 133.8 MB | Download |
| tokenizer-LFM2.5-Audio-1.5B-F32.gguf | GGUF | F32 | 267.6 MB | Download |
| vocoder+tokenizer-LFM2.5-Audio-1.5-BF16.gguf | GGUF | BF16 | 503.0 MB | Download |
| vocoder-LFM2.5-Audio-1.5B-BF16.gguf | GGUF | BF16 | 369.2 MB | Download |
Model Details
Model README
---
base_model: LiquidAI/LFM2.5-Audio-1.5B
base_model_relation: quantized
library_name: gguf
tags:
- gguf
- quantized
---
LFM2.5-Audio-1.5B GGUF
Highest quality quant - original LiquidAI's GGUFs are FP16, which may degrade quality by clipping values. (Or maybe not, but why risk it?)
In this repo:
- Base model in BF16, and Q6_K converted from BF16
- Mmproj in BF16
- Vocoder in BF16
- TTS tokenizer in BF16 and F32 (original tensors are F32)
- Combined vocoder+tokenizer in BF16
Base and mmproj converted using base llama.cpp, tokenizer and vocoder converted using the script from https://github.com/ggml-org/llama.cpp/pull/18641
Text and ASR inference works in baseline llama.cpp. For TTS, use their fork.
Run Auguments/LiquidAI-LFM2.5-Audio-1.5B-GGUF-BF16 with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models