MediaTek-Research/Breeze-ASR-25
Partial fine-tune vs openai/whisper-large-v2. Partial change: 59% of the weights changed (attention, embeddings, mlp, norms), in 32 of 32 layers.
Tensors changed: Embeddings 1/3, Norms 105/324, Attention 448/672, MLP 128/256, other 0/4
2 derivatives checked
- handy-computer/Breeze-ASR-25-gguf: Not comparable yet
- eoleedi/Breeze-ASR-25-mlx: Different architecture
2 other repos hold byte-for-byte identical weight files.