This repository packages the text-generation component from cybermotaz/Qwen3-Omni-30B-A3B-Instruct-NVFP4 in the same flat layout as giangndm/qwen3-30b-omni-text-only.

The model is intended for text-only loading. The model.safetensors.index.json file only exposes model.* and lm_head.* tensors from the NVFP4 Thinker shards. The original Talker and Code2Wav components are not included.

config.json: text-only Qwen3MoeForCausalLM config derived from giangndm/qwen3-30b-omni-text-only, with NVFP4 ModelOpt quantization settings from the source model.
model-00001-of-00005.safetensors through model-00005-of-00005.safetensors: NVFP4 Thinker shard files.
model.safetensors.index.json: filtered index for text-only tensors.
generation_config.json and hf_quant_config.json: source generation and quantization metadata.
Tokenizer files copied from the text-only reference layout.

Source

Text-only layout reference: giangndm/qwen3-30b-omni-text-only
NVFP4 source weights: cybermotaz/Qwen3-Omni-30B-A3B-Instruct-NVFP4

Notes

The shard files are reused from the source Thinker component. The index intentionally omits audio and vision tensors so text-only loaders see the expected causal language model state dict.

giangndm/qwen3-30b-omni-text-only-nvfp4

Qwen3 Omni 30B Text Only NVFP4

Contents

Source

Notes

giangndm/
qwen3-30b-omni-text-only-nvfp4