Extracted using convert_hf_to_gguf.py from llama.cpp. The converter automatically strips thinker. prefix and drops vision/audio/talker/token2wav components, keeping only the text decoder (435 tensors).

yeonseok-zeticai/QWEN_2.5_omni_3b_decoder

Type: Qwen2VL (text decoder)
Parameters: 3.4B (decoder only, excluding vision/audio/talker/token2wav)
Hidden size: 2048
Layers: 36
Attention heads: 16 (KV heads: 2, GQA)
FFN size: 11008
Vocab: 151,936
Context: 32,768 tokens

Qwen2.5-Omni-3B Decoder (GGUF)

Architecture

Files

Usage with llama.cpp

Extraction

yeonseok-zeticai/
QWEN_2.5_omni_3b_decoder