TellWorldAI World Archive
Ask中文
News
Models
Organisations
People
Papers & concepts
Products & agents
Chips & robots
History
Topics
Ask中文

Home›Models

Model

DeepSeek-V3.2 and V3.2-Speciale

DeepSeek-V3.2, released on 1 December 2025, succeeded the experimental V3.2-Exp and integrated thinking into tool use. Its companion V3.2-Speciale, which maximizes reasoning, reached gold-level results at several 2025 competitions.

Status DeprecatedModality TextOpen weights Yes Era Generative AI for everyone

Parameters
685B
License
MIT
Access
API, App, Download
Release date
Developer
DeepSeek

What it is

DeepSeek-V3.2 is a large language model from DeepSeek, released on 1 December 2025 as the official successor to DeepSeek-V3.2-Exp [1]. It went live the same day in DeepSeek's app, on its web chat and through its API [1]. A companion model, V3.2-Speciale, was released alongside it.

Origins

V3.2 follows the experimental V3.2-Exp. The two share an architecture that includes DeepSeek Sparse Attention (DSA), the mechanism DeepSeek introduced to make long-context processing cheaper [2]. The weights of V3.2 total 685 billion parameters and are released under the MIT License [2].

Main features

DeepSeek described V3.2 as its first model to integrate thinking directly into tool use. Tool use means a model calls outside functions, such as search or a code runner, while it works on a task. V3.2 supports tool calls in both thinking mode, in which the model reasons step by step before answering, and non-thinking mode [1]. DeepSeek summed up its aim for the model as 'your daily driver at GPT-5 level performance' [1].

Agent training is the other major focus. DeepSeek said it used a new synthesis method to build agent training data across more than 1,800 environments and 85,000 complex instructions [1]. An agent is a model that carries out multi-step tasks, and training it requires many different simulated settings in which it can practice.

Reported benchmark scores

DeepSeek reported the following scores for V3.2 [2]. On GPQA Diamond, a set of graduate-level science questions, it scored 82.4. On MMLU-Pro, a harder version of a broad knowledge test, it scored 85.0. On SWE-bench Verified, which asks a model to fix real issues from open-source software projects, it scored 70.0, and on the multilingual version of the same benchmark it scored 70.2. On Terminal-Bench 2.0, which tests tasks carried out in a command-line terminal, it scored 39.6.

The V3.2-Speciale companion

V3.2-Speciale is a version designed to maximize reasoning. DeepSeek said it rivals Gemini-3.0-Pro and reached gold-level results at IMO 2025, CMO 2025, the ICPC World Finals and IOI 2025 [1][2]. These are international contests in mathematics, programming and informatics.

Speciale has trade-offs. It needs more tokens to answer, and it does not support tool calls. Until 15 December 2025 it was served only through a temporary API endpoint, at the same price as V3.2 [1]. The final competition submissions for IOI 2025, the ICPC World Finals, IMO 2025 and CMO 2025 were published in the model repository so that others could check them [2].

Weights for both V3.2 and V3.2-Speciale were released on Hugging Face along with a technical report [1].

Position in the series

V3.2 was succeeded by the DeepSeek-V4 preview on 24 April 2026 [4].

CapabilitiesLarge language modelsReasoningAgents & tool use

Ecosystem & regionOpen sourceChina

Version history

12
  1. 2023
    • Nov 29, 2023 DeepSeek LLM (7B and 67B)

      DeepSeek LLM is DeepSeek's first general-purpose language model series, released on 29 November 2023 in 7B and 67B sizes, each with a base version and a chat version.

  2. 2024
    • May 7, 2024 DeepSeek-V2

      DeepSeek-V2 is a mixture-of-experts language model released in May 2024 with 236B total parameters, of which 21B are activated per token. It introduced Multi-head Latent Attention and used DeepSeekMoE for economical sparse training.

    • Sep 5, 2024 DeepSeek-V2.5

      DeepSeek-V2.5 is a model released on 5 September 2024 that merged DeepSeek's chat and code models into one, keeping both conversational and coding ability while improving safety scores.

    • Dec 26, 2024 DeepSeek-V3

      DeepSeek-V3 is an open-weight mixture-of-experts language model that DeepSeek released in December 2024. Each token uses 37 billion of its 671 billion parameters, and DeepSeek describes its chat model as comparable to leading closed-source systems.

  3. 2025
    • Jan 20, 2025 DeepSeek-R1

      DeepSeek-R1 is an open-weight reasoning model that DeepSeek released on 20 January 2025. Built on DeepSeek-V3-Base and trained with reinforcement learning, it is released under the MIT License, which allows commercial use and distillation.

    • Mar 24, 2025 DeepSeek-V3-0324

      DeepSeek-V3-0324 is a March 2025 update to DeepSeek-V3 that keeps the same model structure while raising benchmark scores and moving to the MIT License.

    • May 28, 2025 DeepSeek-R1-0528

      An updated release of DeepSeek's open-weight reasoning model, published on 28 May 2025 under the MIT License, with markedly higher scores on competition math, science and coding benchmarks, plus system-prompt support and stronger function calling.

    • Aug 21, 2025 DeepSeek-V3.1

      DeepSeek-V3.1, released on 21 August 2025 under the MIT License, was DeepSeek's first hybrid model, supporting both a thinking mode and a non-thinking mode in one set of weights. It has 671 billion total parameters, with 37 billion activated per token.

    • Sep 29, 2025 DeepSeek-V3.2-Exp

      An experimental DeepSeek model released on 29 September 2025 that introduced DeepSeek Sparse Attention to make long-context training and inference cheaper, shipped with open weights and an immediate API price cut.

    • Dec 1, 2025 DeepSeek-V3.2 and V3.2-Speciale · This page

      DeepSeek-V3.2, released on 1 December 2025, succeeded the experimental V3.2-Exp and integrated thinking into tool use. Its companion V3.2-Speciale, which maximizes reasoning, reached gold-level results at several 2025 competitions.

  4. 2026
    • Apr 24, 2026 DeepSeek-V4 (V4-Pro and V4-Flash)

      DeepSeek-V4, released in preview on 24 April 2026, comprises V4-Pro and V4-Flash, two open mixture-of-experts models with a 1M-token context. Its architecture targets long-context efficiency, and its later releases extended the series with agent upgrades and multimodal input.

    • Sep 10, 2026 DeepSeek-V4.1-Flash

      DeepSeek-V4.1-Flash, released on 10 September 2026 under the API name deepseek-flash, is a 552-billion-parameter multimodal mixture-of-experts model with a causal encoder-decoder design and a 1M-token context.

20251

  1. ReleaseDeepSeek releases V3.2 and gold-medal-level V3.2-Speciale

    On 1 December 2025 DeepSeek released V3.2, the official successor to V3.2-Exp, and V3.2-Speciale, a version built to maximize reasoning. Both had their weights published on Hugging Face, and V3.2-Speciale was reported to reach gold-level results at several international competitions.

Developed by1

DeepSeek

Succeeded by1

DeepSeek-V4 (V4-Pro and V4-Flash)

Successor of1

DeepSeek-V3.2-Exp

Powers1

DeepSeek app

Uses2

Chain of thought and reasoning modelsMixture of experts (MoE)

Notes and sources

  1. DeepSeek-V3.2 Releasehttps://api-docs.deepseek.com/news/news251201
  2. deepseek-ai/DeepSeek-V3.2https://huggingface.co/deepseek-ai/DeepSeek-V3.2
  3. DeepSeekhttps://en.wikipedia.org/wiki/DeepSeek
  4. DeepSeek V4 Preview Releasehttps://api-docs.deepseek.com/news/news260424

Cite this page: DeepSeek-V3.2 and V3.2-Speciale. TellWorld, https://www.tellworld.net/models/deepseek-v3-2. Updated 2026-11-08