Researchers at Dipsek released a new experimental model called V3.2-ExP on Monday, designed for dramatically low estimated costs when used in contexts for a long time. Deepsek announced the model A post on huggingPosting too A connected academic paper On Github.
The most important feature of the new model is called Deepsek rare meditation, a complex system described in detail in the diagram below. In short, the system uses a module called “Lightning Indier” to prefer specific fractions from the reference window. After that, a separate system that is called “fine token selection system” selects specific tokens from within parts that are to load the limited attention of the module in the window. Together, they allow the rare meditation model to work on longer parts of the context with comparatively small server load.
For long -term reference operation, the benefits of the system are important. Initial testing by Deepsek found that the price of a simple API call could be reduced by half in contexts for a long time. Further testing will be required to create a more strong evaluation, but because the model is open weight and is available to hug independently, it will not be before the third-party tests, can assess the claims made in the paper.
The new model of Deepsac is one of the recent strings of recent successes, which is to deal with the problem of estimates cost-insertately, the server cost of operating a pre-educated AI model, is different from the cost of its training. In the case of Deepsek, the researchers were looking for ways to more efficiently operate the fundamental transformer architecture – and finding that significant improvements are to be made.
Located in China, Deepsek has been an unusual person in AI Boom, especially for those who see AI research as a nationalist conflict between the US and China. The company created waves at the beginning of the year with its R1 model, trained using mainly reinforcement of reinforcement compared to its American competitors. But the model has not made a wholesale revolution in AI training, as some have predicted, and the company has started from the spotlight in the months.
The new “rare meditation” approach is unlikely to produce uproar similar to the R1 – but it can still teach us the providers that some very important tricks to help keep the conclusion cost low.
