DeepSeek-V3.1: A Game-Changing Hybrid AI

DeepSeek has launched a new version of its AI model – DeepSeek-V3.1, which combines the best features of previous models. With improved performance, open source availability, and free chat, it is one of the most exciting AI projects of the year.

Photo: DeepSeek

DeepSeek has unveiled DeepSeek-V3.1, an improved version of its revolutionary artificial intelligence model. The latter was first introduced in December 2024 and immediately ranked among the ten most powerful AI systems in the world. The company surprised the tech world by revealing that it trained the model with significantly fewer computing resources and at a much lower cost than the competition.

The new version works as a hybrid AI model, combining the fast response of the V3 model with the deep thinking of the R1 model. The DeepSeek-V3.1 model is available for free under the MIT open source license. To use the full version of 671B, at least 720 GB of space is required (or 170 GB for the 1-bit quantized version). The smallest model requires a powerful graphics card with at least 24 GB of memory. For example, an Nvidia 5090 with 32 GB, which costs about 1,600 euros.

The DeepSeek-V3.1 model has been improved in programming and surpasses the performance of the previous V3 and R1 models after testing in the SWE-bench benchmark program. It also achieves better results in other thinking tests such as xbench-DeepSearch, SimpleQA and FRAMES AI. With a 128K context window, it allows free chat with AI. So everything that advanced people need.