logic65/Qwen3.6-Whittle-25B-A3B
LoRA-style vs Qwen/Qwen3.6-35B-A3B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, shared_expert_gate, up_proj changed, in 40 of 40 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/1, Attention 0/280, Norms 0/131, MLP 160/160, lm_head 0/1
2 derivatives checked
- mradermacher/Qwen3.6-Whittle-25B-A3B-GGUF: Quantization
- logic65/Whittle-Next-26B-A3B: Not comparable yet