d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE

Full fine-tune vs ukisai/Swift-1.5-Qwen3.8-Flash-Next-NVFP4. Fine-tune: 90% of the weights changed, across 1 of 48 layers.

Tensors changed: Embeddings 128/137, MLP 0/1943, Attention 0/608, Norms 0/302, other 0/8, lm_head 0/1

1 derivatives checked

• d0xin/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-FP8PLE-Rank2-Abliterated: Partial fine-tune

On the Hub · Downloads on Model Pulse