Skip to content

DeepSeek V3.2

by DeepSeek · deepseek-v3 family

671B

parameters

text-generation code-generation reasoning multilingual tool-use math

DeepSeek V3.2 is the latest update to DeepSeek's flagship MoE model, with 671B total parameters and 37B active per token. It features DeepSeek Sparse Attention (DSA) for improved long-context efficiency and enhanced reasoning capabilities. Like its predecessor, V3.2 requires enterprise-class hardware for local deployment. At Q4 quantization it needs ~420 GB of VRAM, putting it in multi-GPU or very high-memory Mac territory. A cloud variant is also available through Ollama.

Quick Start with Ollama

ollama run q4_K_M
Resources Ollama Hugging Face Official Page
Creator DeepSeek
Parameters 671B
Architecture transformer-decoder
Context 128K tokens
Released Mar 1, 2026
License DeepSeek License
Ollama deepseek-v3.2

Quantization Options

Format File Size VRAM Required Quality Ollama Tag
Q4_K_M rec 400 GB 420 GB q4_K_M

Compatible Hardware

Q4_K_M requires 420 GB VRAM

Compatible Hardware

HardwareVRAMTypeFitEst. Speed
Mac Studio M4 Ultra 512GB512 GBmacRuns~2 tok/s
106 hardware device(s) cannot run this model at Q4_K_M.