On Monday, researchers at DeepSeek released a new experimental model called V3.2-EXP, designed to dramatically reduce inference costs when used in long context operations. Deepseek announced the model in a post about Face’s hugs and posted an academic paper linked to Github.
The most important feature of the new model is called DeepSeek Sparse Anterest. This is a complex system explained in detail in the diagram below. Essentially, the system uses a module called “Lightning Indencer” to prioritize certain excerpts from the context window. Another system, called the “fine-grained token selection system,” then selects a specific token from within these excerpts and loads it into the module’s limited attention window. In summary, sparse attention models can work so that server loads over long sections of relatively small contexts.
For long-context operations, the advantages of the system are important. A preliminary test by DeepSeek shows that the price of simple API calls can be reduced by half in long context situations. Building a more robust assessment will require further testing, but since the models are openweight and freely available, it will not be long before third-party tests can evaluate claims made in the paper.
Deepseek’s new model is one of the recent breakthroughs tackling the issue of inference costs. Essentially, it is the server cost for manipulating a pre-trained AI model that is different from the cost of training. In Deepseek’s case, researchers were looking for ways to make basic transformer architectures work more efficiently.
China-based Deepseek was a rare figure in the AI boom, especially those who view AI research as a nationalist struggle between the US and China. The company made waves in the R1 model early in the year, and was trained using reinforcement learning, primarily at a much lower cost than its American competitors. However, this model has not triggered a wholesale revolution in AI training, as some have predicted. The company then retreated from the spotlight in those few months.
The new “sparse attention” approach is unlikely to produce the same uproar as R1, but it can teach providers the tricks needed to keep inference costs low.
