DeepSeek-V3.2

Updated
25.06.2026
Thinking
Tools
Reasoning
Code
Multilingual
Run

huggingface-cli download deepseek-ai/DeepSeek-V3.2
from transformers import AutoModel
model = AutoModel.from_pretrained("deepseek-ai/DeepSeek-V3.2")
More models
View all

  
  

NameSize / UsageContextInput
DeepSeek-V4-Pro
1MText
DeepSeek-V4-Flash
1MText
DeepSeek-R1
128KText
At a glance
  • License: Mit
  • Context length: 128K tokens
  • Languages: Multilingual
  • Minimum hardware: ~380 GB VRAM
  • Strengths: reasoning, agentic coding and long context
Overview
DeepSeek-V3.2 is a 685.4B-parameter large language model from deepseek-ai, built on a Mixture-of-Experts (MoE) design so only a fraction of those parameters fire on any given token. It pairs that MoE backbone with DeepSeek Sparse Attention (DSA), an attention mechanism that cuts the compute cost of long context while keeping quality high across its 128K-token window. The model handles a reasoning ("thinking") mode, tool calls, code, and multilingual input.
In Atomic Chat the model runs on your own machine. Weights load locally, prompts never leave the device, and once the files are downloaded it works offline with no API key and no per-token billing.
What it is good at
DeepSeek-V3.2 leans on its reasoning, tool-use, code, and multilingual capabilities. Three things it does well:
  • Step-by-step reasoning — the thinking mode works through math, logic, and multi-stage problems before answering, which suits analysis and planning tasks.
  • Tool-using agents — it calls external tools in both thinking and non-thinking modes, so you can wire it into local scripts, search, or function-calling workflows.
  • Code and multilingual work — it generates and explains code and handles prompts across many languages, useful for development and cross-language drafting or translation.
Running it locally
At 685.4B parameters this is a large model. Unquantized FP16 weights run into the hundreds of gigabytes; 4-bit quantization brings the footprint down toward roughly 386 GB, still aimed at multi-GPU or high-memory workstations rather than a typical laptop. The 128K context window also adds to memory use as conversations grow. Pull the weights from Hugging Face:
huggingface-cli download deepseek-ai/DeepSeek-V3.2
From there you can serve it with vLLM or load it through Transformers, or open it in Atomic Chat with one click and let the app handle the local setup.
License
DeepSeek-V3.2 is released under the MIT license. That covers both code and weights, and it permits local deployment, modification, fine-tuning, distillation, and commercial use with no royalties or revenue share.
Desktop



macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download
Frequently asked questions
DeepSeek-V3.2 is a 685.4B-parameter open-weight language model from deepseek-ai. It uses a Mixture-of-Experts design with DeepSeek Sparse Attention for efficient long-context work, and supports a thinking mode, tool calls, code, and multilingual input across a 128K-token window.
It is a large model, so plan for serious memory. Full FP16 weights run into the hundreds of gigabytes, and even 4-bit quantization needs roughly 386 GB, which points to multi-GPU or high-memory workstation setups rather than a single consumer card. The 128K context window adds memory pressure as the conversation grows.
Yes. The weights are published on Hugging Face under the MIT license at no cost. You can download and run it without an API key or per-token fees, and the license allows both personal and commercial use.
Yes. Once the weights are downloaded, DeepSeek-V3.2 runs fully on your own hardware with no internet connection required. In Atomic Chat your prompts stay on the device, which suits private or data-sensitive work.
Download the weights with huggingface-cli download deepseek-ai/DeepSeek-V3.2, then serve them with vLLM or Transformers, or open the model in Atomic Chat with one click. It is a strong fit for step-by-step reasoning, tool-using agents, and coding tasks, and it handles multilingual prompts.