DeepSeek오픈소스

DeepSeek V3.2 (正式版)

이 모델 비교

DeepSeek-AI의 파운데이션 모델입니다.

파라미터

Undisclosed

컨텍스트

라이선스

https://github.com/deepseek-ai/deepseek-LLM/blob/main/LICENSE-MODEL

출시일

2025-12-01

벤치마크 성능

AA Intelligence Index

LMArena Elo

HLE

ARC-AGI-2

SWE-bench Verified

GPQA Diamond

MMLU-Pro

LiveCodeBench

AIME 2025

MATH-500

일본어 처리 능력

High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API 가격

입력 가격 (1M 토큰당)

$0.28

출력 가격 (1M 토큰당)

$

과금 모드: standard

강점

    약점

      활용 사례

        심층 분석

        Arena Elo

        1425

        #56 overall on BenchLM provisional leaderboard

        SWE-Bench Verified

        73.10%

        non-thinking mode

        GPQA Diamond

        82.40%

        thinking mode

        Input Price

        $0.28/1M tokens

        ~1/10 of GPT-5's price

        Context Window

        128K tokens

        suitable for most production tasks

        Parameters

        671B total (37B active)

        efficient Mixture-of-Experts architecture

        강점

        • Exceptional cost-performance ratio, especially for Chinese language tasks.
        • First model to integrate chain-of-thought reasoning with tool use, enhancing agent capabilities.
        • Open-source under MIT license, enabling self-hosting and customization.

        약점

        • Text-only model; no multimodal (image, video) support.
        • Long-context handling degrades near the 128K limit (lost-in-the-middle phenomenon).
        • Agent tool use in long loops (10+ rounds) is less reliable compared to top closed-source models like Claude.

        경쟁사 비교

        ModelArenaSWEGPQAPrice
        DeepSeek V3.2142573.10%82.40%$0.28/$0.42
        Claude Opus 4.7N/AN/A87.30%$15/$75

        DeepSeek V3.2 is a 671-billion-parameter Mixture-of-Experts foundation model released by DeepSeek-AI in December 2025. It is the official production version of the DeepSeek V3.2 series, designed to balance strong reasoning capabilities with practical output length for general-purpose use. The model introduces DeepSeek Sparse Attention (DSA) for improved long-context efficiency and is the first from DeepSeek to integrate chain-of-thought reasoning with tool use, enabling more sophisticated agent workflows.

        Positioned as a high-value, open-weight alternative to frontier closed-source models like GPT-5, DeepSeek V3.2 delivers competitive performance across reasoning, coding, and agent benchmarks at a fraction of the cost. Its extensive reinforcement learning training on a massive synthetic agent task dataset (1800+ environments, 85,000+ tasks) aims to enhance real-world generalization. While it lacks multimodal support and has some limitations in very long-context stability and complex multi-step agent loops, it represents a significant step in making advanced AI capabilities accessible and affordable.

        분석 생성일: 2026-07-17