Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B
LoRA-style vs unsloth/Qwen3.5-9B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, k_proj, o_proj, out_proj, q_proj, up_proj, v_proj changed, in 32 of 32 layers; norms, biases and embeddings are untouched.
Tensors changed: lm_head 0/1, Embeddings 0/5, MLP 96/211, Attention 56/336, other 0/1, Norms 0/221
6 derivatives checked
- mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-GGUF: Same weights, converted
- mradermacher/DeepSeek-V4-Pro-Qwen3.5-9B-i1-GGUF: Quantization
- mlx-community/DeepSeek-V4-Pro-Qwen3.5-9B-4bit: Quantization
- mlx-community/DeepSeek-V4-Pro-Qwen3.5-9B-bf16: Partial fine-tune
- mlx-community/DeepSeek-V4-Pro-Qwen3.5-9B-8bit: Quantization
- mlx-community/DeepSeek-V4-Pro-Qwen3.5-9B-6bit: Quantization