Everything about DeepSeek V3.2

Overview

DeepSeek released DeepSeek V3.2 on December 1, 2025. It has a reasoning mode: it can "think" before answering, which helps with multi-step problems but uses more tokens, and therefore more money. Its weights are published: you can download it and run it on your own servers, or use one of the 31 API providers that list it on models.dev. That makes it 10 months old as of October 22, 2026.

We file it under open-weight models (68 models in this category on the site).

Benchmark results

According to results published by Epoch AI (checked on October 22, 2026), DeepSeek V3.2 has a capability index of 59.8 out of 100, computed on 4 benchmarks from our basket.

It is not among the top 30 of the weekly ranking.

• GPQA Diamond (PhD-level science questions): 83.4% — the best tracked model, Gemini 4 Argon, reaches 96.4%
• OTIS Mock AIME (competition mathematics problems): 87.8% — the best tracked model, GPT-5.5 Pro, reaches 100%
• Terminal-Bench (agent tasks in a terminal): 39.6% — the best tracked model, GPT-5.5, reaches 84.7%
• APEX-Agents (professional tasks carried out by an agent): 21.3% — the best tracked model, Gemini 4 Argon, reaches 82.4%

How much does DeepSeek V3.2 cost?

DeepSeek V3.2 costs $0.29 per million input tokens and $0.43 per million output tokens (median of 31 providers, checked on October 22, 2026).

To put it in perspective: reading 20,000 tokens (about thirty pages) and writing 2,000 costs roughly $0.007; a thousand of those tasks, $6.66.

That is 2.3 times cheaper than the median output price of open-weight models ($1).

To estimate your own budget, use the agent cost calculator.

Technical specifications

Its context window reaches 128,000 tokens, roughly 192 pages of text at 500 words per page: enough for a medium-length document. A single answer can run to 128,000 tokens.

It accepts as input: text. Its knowledge stops at December 2024: beyond that, it needs web search or your documents.

It can call tools (functions, APIs), which makes it a possible foundation for an AI agent.

The weights are published on Hugging Face (deepseek-ai/DeepSeek-V3.2): 685.4 billion parameters, mit license and 5,461,313 downloads over the last 30 days. See also our list of open-source models.

When to choose it

Given its category, DeepSeek V3.2 is mainly aimed at projects where you want control over your data or hosting, or want to adapt the model to your field.

Before adopting it, test it on your own examples: benchmarks measure standard tasks, not your job.

What could replace it?

In the same category, the models closest to DeepSeek V3.2 are:

• Qwen3.5 397B-A17B (Alibaba Cloud (Qwen), index 59.2)
• MiniMax-M3 (MiniMax, index 55.5)
• Qwen3.6 35B-A3B (Alibaba Cloud (Qwen), index 55.3)
• Kimi K2.7 Code (Moonshot AI, index 66.7)