deepseek-ai/DeepSeek-V3.2

Full fine-tune vs deepseek-ai/DeepSeek-V3.2-Exp-Base. Fine-tune: 100% of the weights changed, across 62 of 62 layers.

Tensors changed: Embeddings 2/2, Attention 780/930, Norms 375/376, MLP 1518/1689, lm_head 1/1, other 2/2

14 derivatives checked

• unsloth/DeepSeek-V3.2-GGUF: Quantization
• nvidia/DeepSeek-V3.2-NVFP4: Different architecture
• DevQuasar/deepseek-ai.DeepSeek-V3.2-GGUF: Quantization
• amd/DeepSeek-V3.2-mxfp4: Different architecture
• mlx-community/DeepSeek-V3.2-4bit: Different architecture
• mlx-community/DeepSeek-V3.2-8bit: Different architecture
• cerebras/DeepSeek-V3.2-REAP-345B-A37B: Partial fine-tune (fewer layers)
• cerebras/DeepSeek-V3.2-REAP-508B-A37B: Partial fine-tune
• QuantTrio/DeepSeek-V3.2-AWQ: Different architecture
• lovedheart/DeepSeek-V3.2-REAP-345B-A37B-GGUF-Experimental: Quantization
• mlx-community/DeepSeek-V3.2_bf16: Quantization
• amd/DeepSeek-V3.2-mtp-ptpc: LoRA-style
• RedHatAI/DeepSeek-V3.2-NVFP4-FP8-BLOCK: Different architecture
• baa-ai/DeepSeek-V3.2-RAM-350GB-MLX: Different architecture

On the Hub · Downloads on Model Pulse