DeepSeek-V3.2-4L

For testing serving infrastructure only. Its output is not useful text.

deepseek-ai/DeepSeek-V3.2 (revision a7e62ac04ecb2c0a54d736dc46601c5606cf10a6) cut to its first 4 of 61 decoder layers and the MTP layer 61 (renumbered 4). Every kept tensor is copied byte for byte, so the quantization, the attention and DSA indexer shapes, the KV cache layout, the MoE routing and the speculative-decoding layer are those of the full model; only the depth differs. 32.5 GB.

Built by cut.py with:

cut.py --source deepseek-ai/DeepSeek-V3.2 --revision a7e62ac04ecb2c0a54d736dc46601c5606cf10a6 --layers 4 --repo nvcnvn/DeepSeek-V3.2-4L --license-from deepseek-ai/DeepSeek-V3.2 --out /home/runner/work/_temp/out

The original license is in LICENSE.

Downloads last month: 34

Safetensors

Model size

29B params

Tensor type

BF16

F8_E4M3

F32

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for nvcnvn/DeepSeek-V3.2-4L

Base model

deepseek-ai/DeepSeek-V3.2-Exp-Base

Finetuned

deepseek-ai/DeepSeek-V3.2

Quantized

(20)

this model