DeepSeek has announced V3.1, an upgrade to its large language model. The release took place on 19 August 2025 through the company’s official WeChat group. Though the announcement was low-key, the AI community has been quick to react. Developers and researchers are calling it a step forward for open-source models.
The most notable improvement is the expanded context length. V3.1 can now handle 128,000 tokens in a single query. This matches the open-source version and allows the model to manage long conversations, technical documents, and retrieval-based tasks with better Accuracy. Enterprises see this as a strong feature for data-heavy workflows.
V3.1 also raises the parameter count to 685 billion, compared to 671 billion in V3. Despite the increase, costs remain under control thanks to its Mixture-of-Experts design. Only 37 billion parameters are active per token, which reduces the expense of running the model. This efficiency makes it competitive with closed systems such as GPT-4o and Claude 3.5 Sonnet.
Stronger in coding, logic and math
Community testing has shown that V3.1 performs better in problem-solving. It has been able to complete tasks involving complex rules and logical reasoning. Developers also note stronger results in coding, especially in Python and Bash. Accuracy benchmarks now stand close to 60 per cent, which is higher than before.
Mathematics is another clear improvement. The model builds on V3’s success, which outperformed rivals like Qwen2.5 72B on tests such as AIME and MATH-500. These results confirm V3.1’s value for users who work on scientific or analytical projects.
Open-Source release and future outlook
DeepSeek has continued its open approach by releasing V3.1 under the MIT Licence. Developers can access the model on Hugging Face in Safetensors format. While major inference providers have yet to add support, the release is already in use within open-source communities.
Training costs were only 5.6 million US dollars, achieved with 2.788 million H800 GPU hours. In comparison, proprietary models often cost more than one hundred million to build. This cost advantage has earned DeepSeek the title “the Pinduoduo of AI,” showing its ability to deliver at scale without huge budgets.
V3.1 works smoothly with existing APIs, making integration simple for businesses. It is available through the company’s website, mobile app, and WeChat mini-program. The knowledge cut-off stands at July 2025. Online forums have started speculating that V3.1 could be followed by DeepSeek-R2, a reasoning-focused release expected in 2026. DeepSeek V3.1 is already being viewed as more than a routine update. It signals a closing gap between open and closed systems. With stronger reasoning, more context, and reduced costs, V3.1 positions DeepSeek as a serious player in the AI race.