logic65/Qwen3.8-Whittle-MoE-27B-A17.8B
Partial fine-tune vs Qwen/Qwen3.8-27B. Partial change: 74% of the weights changed (attention, norms), in 64 of 64 layers.
Tensors changed: Norms 143/209, Attention 304/448, Embeddings 0/1, lm_head 0/1
4 derivatives checked
- logic65/Qwen3.8-Whittle-MoE-27B-A17.8B-GGUF: Quantization
- mradermacher/Qwen3.8-Whittle-MoE-27B-A17.8B-GGUF: Full fine-tune
- scima/Qwen3.8-Whittle-MoE-27B-A17.8B-GGUF-all-quants: LoRA-style
- barozp/Qwen3.8-Whittle-MoE-27B-MLX-4bit: Quantization
2 other repos hold byte-for-byte identical weight files.