DeepSeek, one of the leading innovators in the Chinese AI industry, recently revealed a groundbreaking new AI technology. Having recently finalized their new experimental model, V3.2-exp, this new, groundbreaking model further cuts inference costs substantially, making it feasible for long-context operations. It showcases a big step forward in the efficiency of the basic transformer architecture.
Engineers at DeepSeek concentrated on improving the transformer architecture to run as efficiently as possible, and they made some powerful strides with V3.2-exp. The model uses their new system DeepSeek Sparse Attention, which are yielding strong preliminary results. During playtesting, V3.2-exp proved to be surprisingly effective. Second, it significantly decreases API call costs—up to 50% of the cost—when working with large contexts.
DeepSeek has made V3.2-exp available with open weights on Hugging Face. This joint effort is focused on encouraging cooperation and building support for additional research. This move reinforces the company’s dedication to driving AI technology forward while engaging communities in the process. The company shared a linked academic paper detailing their findings on GitHub, providing transparency and encouraging further exploration in the field.
DeepSeek shared their process and detailed findings in a post on HuggingFace. They highlighted the importance of these findings and how they could move the industry forward. V3.2-exp which just released, takes that success from their previous model – R1 – and expands on it. R1 made quite a splash earlier this year with its relatively inexpensive reinforcement learning based training, beating out most American models at significantly less cost.
DeepSeek is no typical David among AI industry Goliaths. They are repeatedly able to hit breakthroughs that further reduce inference costs. Their new “sparse attention” method yields some fascinating knowledge. U.S. providers should heed these lessons to maintain low operational costs while increasing performance under the model.
With V3.2-exp DeepSeek is at the cutting edge of AI efficiency with smart databases. This would be a game changer in how long-context operations are handled across a wide variety of applications.
“lightning indexer” – source not explicitly mentioned
“fine-grained token selection system” – source not explicitly mentioned