Archive · 20 Aug 2025
DeepSeek V3.1 Ushers in Advancements in Open-Source AI
DeepSeek, an AI company, has released a new, improved version of its AI model called V3.1.
This new version can understand much longer texts, like long articles or technical documents, and is better at answering questions that need reasoning or problem-solving skills.
Even though the new version has more parameters, it is still relatively cheap to run thanks to its special design.
The company also made the new AI available for free to use, which allows developers to improve it and build new projects.
DeepSeek has released V3.1, an upgraded large language model, on August 19, 2025.
V3.1 features an expanded context length, now handling 128,000 tokens per query.
The model's parameter count is 685 billion, with a Mixture-of-Experts design to control costs.
V3.1 shows improved performance in coding, logic, math, and problem-solving.
The open-source model is available on Hugging Face under the MIT Licence.
- Who
- DeepSeek developers and the AI community.
- What
- The release and features of DeepSeek's V3.1 large language model.
- When
- Announced on August 19, 2025.
- Where
- Open-source communities and AI platforms.
- Why
- To provide an improved, open-source large language model.
Developers
Proprietary Model Providers
Open Source vs. Proprietary Models
Developers
Benefit from open-source access.
Proprietary Model Providers
Benefit from proprietary model advantages.
- Model
- DeepSeek V3.1
- Release Date
- August 19, 2025
- Context Length
- 128,000 tokens
- Parameter Count
- 685 billion
- License
- MIT
- Training Cost
- 5.6 million USD
Sources
Sam Altman admits China’s DeepSeek forced OpenAI’s hand on open models: ‘If we didn’t do it…’
OpenAI CEO Sam Altman warns America is underestimating China’s AI progress
OpenAI’s Sam Altman says AI market bubble is real but also sees long-term benefits
DeepSeek-R1: Hype cools as India seeks practical GenAI solutions