SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

DeepSeek V3.1 - DeepSeek's latest open-source AI model version

Author: AI小集 · Original source: AI工具集

What is DeepSeek V3.1

DeepSeek V3.1 isDeepSeekthe latest AI model version released, featuring a hybrid reasoning architecture that can freely switch between thinking mode and non-thinking mode, with significantly improved thinking efficiency. The model has undergone multiple upgrades on the basis ofV3, with the context window expanded from 64k to 128k, enabling it to process longer text inputs. The model uses a Mixture of Experts (MoE) architecture, with the same number of parameters as V3. The model performs excellently in coding and search agents, and can efficiently complete complex tasks. The model is now available onDeepSeek web version、App, and the models called by the mini program and the API open platform have all been updated simultaneously. The new model's self-awareness is DeepSeek V3.

DeepSeek V3.1

Main features of DeepSeek V3.1

  • Natural language processing: DeepSeek V3.1 can generate high-quality text, such as creative writing and story creation, and its tone when answering questions is more lively and information-rich.
  • Coding ability: The model can generate more complex and more complete code.
  • Multi-domain knowledge: Answers to niche historical questions, etc. are more accurate and more informative, and it can also provide in-depth answers in the fields of technology and science.
  • Hybrid reasoning architecture: Supports free switching between thinking mode and non-thinking mode, adapting to the needs of different scenarios.
  • Efficient thinking ability: Compared with the previous generation, thinking efficiency is higher, output is more concise, and performance is on par.
  • Powerful Agent capabilities: Performs excellently in coding agent and search agent tasks, and can complete complex code fixes and multi-step reasoning searches.
  • API feature upgrades: Supports deepseek-chat and deepseek-reasoner two modes, the context is expanded to 128K, and the Anthropic API format is supported.
DeepSeek V3.1

DeepSeek V3.1 project address

  • HuggingFace model library：
    • Base model：https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base
    • Post-trained model：https://huggingface.co/deepseek-ai/DeepSeek-V3.1

How to use DeepSeek V3.1

  • Web version: VisitDeepSeekthe official website and use the latest version of the model.
  • App: Download and installDeepSeek App, log in to your account, and then use the V3.1 model.
  • Mini program: Search for the DeepSeek mini program on WeChat or other platforms, and use it after logging in to your account.
  • API open platform: For API calls, you need to first register and obtain an API key, and make calls according to the guidelines in the official documentation. The API supports thinking mode (deepseek-reasoner) and non-thinking mode (deepseek-chat), the context window is expanded to 128K, and it supports Function Calling instrictmode.
  • Visit the Hugging Face platform: Visit the DeepSeek-V3.1-Base model page on Hugging Facehttps://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base
    • Model download: Download the model weights, configuration files, and script code.
    • Local deployment: Install the necessary dependencies in the local environment (such as Python, the Transformers library, etc.), load the model, and perform inference.
    • API calls: Use the API interface provided by Hugging Face to integrate the DeepSeek-V3.1 model into your own application or service, enabling automated processing and interaction.

DeepSeek V3.1 API pricing

API call pricing will be adjusted starting in the early hours of September 6, 2025, and the nighttime discount will be canceled.

  • Input price：
    • Cache hit: 0.5 yuan per million tokens
    • Cache miss: 4 yuan per million tokens
  • Output price: 12 yuan per million tokens

Application scenarios of DeepSeek V3.1

  • Content creation: Used to generate creative text such as articles, stories, and poems, helping creators spark inspiration and improve creative efficiency.
  • Programming assistance: Helps developers quickly generate code frameworks, optimize code logic, and improve programming efficiency, suitable for front-end development and mini-game development.
  • Education: As a teaching assistance tool, it provides students with knowledge answers and explains complex scientific and historical questions.
  • Scientific research: Assists researchers in organizing and analyzing data, and provides answers to scientific questions and analytical approaches.