microsoft/VibeVoice-Realtime-0.5B
Full fine-tune vs Qwen/Qwen2.5-0.5B. Fine-tune: 100% of the weights changed, across 20 of 20 layers.
Tensors changed: Embeddings 1/1, Norms 41/41, MLP 60/60, Attention 140/140
2 derivatives checked
- cstr/vibevoice-realtime-0.5b-GGUF: Not comparable yet
- mohammed-bahumaish/vibevoice-realtime-0.5b-with-encoder: Same weights, repacked (fewer layers)
13 other repos hold byte-for-byte identical weight files.