What is the difference between "chat" and "reasoner"?

deepseek-chat is the standard mode (V3.2 Non-thinking) for fast answers. deepseek-reasoner activates the "Thinking" mode (V3.2 Thinking), where the model pauses to generate a hidden chain of thought before answering, perfect for hard math problems.

Can I run DeepSeek V3.2 locally?

You can run the smaller distilled versions (like 7B or 8B) on a consumer GPU. However, the full V3.2 671B model requires massive VRAM (multiple A100/H100 GPUs), making it practical only for data centers.

How much does it cost?

The API pricing is $0.28 per million input tokens (or $0.028 if cached!) and $0.42 per million output tokens. This is roughly 10-20x cheaper than GPT-4o.

DeepSeek V3.2

Frontier intelligence at a fraction of the cost.

From DeepSeek AI

Founder: DeepSeek AI Team (Liang Wenfeng, Founder of High-Flyer Quant)

Text Generator

Sep 2025

38 Million+

Freemium

Pricing Model

Freemium

Starting Price

N/A

About This AI

DeepSeek V3.2 is the latest flagship open-source model family from the Chinese research lab DeepSeek AI, known for shattering the “price-performance” barrier.

Released in late 2025, V3.2 introduces a unified “Hybrid Reasoning” architecture that allows a single model to switch seamlessly between a standard fast chat mode and a slow, deliberative “Thinking Mode” (similar to OpenAI o1).

It utilizes a novel “DeepSeek Sparse Attention” (DSA) mechanism to drastically reduce inference costs, making it the most affordable frontier-class model on the market. The release also includes “V3.2-Speciale,” a variant specifically tuned for solving complex math and coding problems that stump even GPT-5 level models.

AI Art Chatbot Coding LLM Text To Image

Pricing

Pricing Model

Freemium

Starting Price

N/A

Key Features

Hybrid Reasoning Modes: Users can toggle "Thinking Mode" on or off within the same model API (deepseek-reasoner vs deepseek-chat), enabling deep Chain-of-Thought (CoT) only when needed.
DeepSeek Sparse Attention (DSA): A breakthrough attention mechanism that focuses only on relevant tokens, reducing computational cost by over 50% for long contexts.
V3.2-Speciale: A specialized high-compute variant designed purely for deep reasoning tasks (Math, Physics, Coding) that surpasses GPT-5 in specific benchmarks.
Multi-Token Prediction (MTP): Predicts multiple future tokens simultaneously during generation to speed up inference without losing quality.
Auxiliary-Loss-Free Load Balancing: A unique Mixture-of-Experts (MoE) strategy that ensures all model "experts" are used efficiently without the performance penalty of traditional balancing losses.
Massive Scale: 671 Billion total parameters with only 37 Billion active per token, offering flagship intelligence with the speed of a smaller model.

Pros

Unbeatable Price: At ~$0.28 per million input tokens, it is significantly cheaper than Western competitors like Claude 3.5 or GPT-4o.
Open Weights: The base models are open-source (MIT License), allowing for unrestricted local use and fine-tuning.
"Thinking" Capability: One of the few open models that natively supports "System 2" reasoning chains.
Context Caching: Features a "Context Caching" price of just $0.028/1M tokens, making repetitive tasks (like chatting with a codebase) nearly free.
Strong Coding: Comparable to Claude 3.5 Sonnet and GPT-5.1 Codex in coding benchmarks.

Cons

Hardware Heavy: Running the full 671B model locally is impossible for most users; it requires massive multi-GPU clusters (H800s).
Speciale Limitations: The "Speciale" variant does not support tool calling or functions, limiting it to pure text reasoning.
Chinese-Bias: While excellent in English, it occasionally defaults to Chinese cultural references or language nuances in ambiguous situations.

Best for Developers needing cheap but powerful API access, researchers studying MoE architectures, and enterprises looking to self-host high-intelligence models to avoid data privacy issues.

FAQs

What is the difference between “chat” and “reasoner”?
deepseek-chat is the standard mode (V3.2 Non-thinking) for fast answers. deepseek-reasoner activates the “Thinking” mode (V3.2 Thinking), where the model pauses to generate a hidden chain of thought before answering, perfect for hard math problems.
Can I run DeepSeek V3.2 locally?
You can run the smaller distilled versions (like 7B or 8B) on a consumer GPU. However, the full V3.2 671B model requires massive VRAM (multiple A100/H100 GPUs), making it practical only for data centers.
How much does it cost?
The API pricing is $0.28 per million input tokens (or $0.028 if cached!) and $0.42 per million output tokens. This is roughly 10-20x cheaper than GPT-4o.