When the Depseek chatbot was presented a few months ago, the world of AI trembled. Chinese chatbot is not alone Open Source, but it is also freeHe trains more easily and with an infinitely more modest hardware than his main rivals (chatgpt, twins and llama). This made it popular in a short time, although mainly among AI fans.
Since then, the model has improved in some things through different minor updates, but now it has just made an official leap towards a new version. Deepseek V3.1 is already a reality and we will show you its news, Are you prepared?
Deepseek V3.1 can analyze books up to 400 pages and everything works in a single unified model
Through a modest publication in WeChat, on August 19, the company behind Deepseek announced that its model officially reached version 3.1. The main novelty is The expansion of the context window, which is now 128,000 tokenIn addition to increasing the number of parameters to 685,000 million.
With this, Deepseek V3.1 May Manage much more extensive data inputs and outputsequivalent to a book from 300 to 400 pages. This is particularly useful when Analyze the technical documents and when deep research is conductedIn addition to helping to contain extensive conversations without losing the thread. According to developers, Deepseek V3 was already compatible internally with this expanded context, but right now that it was officially enabled for different tools.
Next to this, Depseek has eliminated all references to the R1 model In the chatbot interface, which advances that the transition to a unique hybrid model is already a reality. Said in a simpler way: in Deepseek V3.1 there is no longer a reasoning model and another for activities without reasoning, but Everything has been integrated into an excellent unified model.
Las First performance tests They also announced, because Deepseek V3.1 obtained a Score 71.6% in the Aider programming test. This places him above Claude Opus 4, making sure that Deepseek is considered as One of the most powerful Open Source models Currently available. There have also been improvements in terms of logical and mathematical reasoning, although not so large if compared with R1-0528, the previous version.
As for the implementation of your bees in other tools or the collection of your request to use offline, Deepseek V3.1 can operate in multiple precision formats, such as previous versions. It can work in the BF16, FP8 and F32 modeThus giving greater flexibility to users, depending on the environment in which it will be implemented. The source code It can be downloaded via embraceAnd it has an open source license approved by MIT.
The public expected the launch of R2, but Depseek is not yet ready for that update
Although V3.1 is an update that many will appreciate, the most faithful users in Deepseek They expected the next great launch to be the R2 model. This would arrive with real Improvements in CLM reasoning skillsbut a Report published in Financial Times It indicates the reason why it has not yet arrived.
According to the document, the company behind Depseek is having Huawei ascended chips problems. The company wants to reduce its dependence on Nvidia and align the Chinese strategy of being self -sufficient in terms of artificial intelligence. However, Ascend Chips Training did not give good resultsespecially for compatibility problems.
Deepseek has decided to go on with the Nvidia chips for a little more time, so it was launched V3.1. R2 will continue on the Road Map, with axes as the main hardwareBut It will not be launched until the problems are solved.
