DeepSeek open-sources tool to help cut AI operating costs

DeepSeek, the Chinese AI lab founded by entrepreneur Liang Wenfeng, has open-sourced DeepInfer, the software that powers its own models. The tool makes the company’s AI models run faster and use computing resources more efficiently.

“We’re sharing our production stack (the software that powers the models – Oninvest) to help the community lower compute costs, reduce energy use, and make AI more affordable,” the company stated on X on Tuesday.

DeepInfer currently supports only DeepSeek’s V4.1-Flash AI model, but the company stated that support for future models would follow, Bloomberg reports. The tool is designed to reduce the cost of inference, the stage at which a trained model generates responses. These costs play a major role in the economics of AI services: training a model involves a large, one-time expense, while inference incurs a cost with every user request.

The release of DeepInfer’s source code offers a glimpse into technical methods that AI companies typically keep secret and shows how DeepSeek keeps inference costs low, Bloomberg explains. DeepInfer combines different techniques, such as distributed inference, which spreads workloads across different hardware. These methods allow models to generate more responses in less time and get more performance out of chips. Many of the techniques are widely used, but DeepSeek has tailored them to work efficiently with its models, Bloomberg notes.

More efficient software can lower the cost of processing each request and potentially reduce energy use, Bloomberg points out. This could improve margins for cloud and hosting providers deploying DeepSeek’s models and lead to lower prices for end customers. For companies running DeepSeek’s models on their own infrastructure, DeepInfer could help lower operating costs.

There are no independent benchmarks yet confirming DeepInfer’s efficiency gains, Bloomberg cautions. However, while the tool is primarily designed for DeepSeek’s models, other developers can also study the open-source code and adapt its techniques for their own inference systems.

DeepSeek rose to global prominence in early 2025 after releasing its R1 model. The company claimed that V3, the foundation model underlying R1, had cost less than $6 million to train, substantially less than comparable U.S. models. DeepSeek’s low-cost models triggered a $1 trillion selloff in tech stocks amid concerns that the AI industry was overspending on computing power.

DeepSeek is also developing its own inference chip to reduce its reliance on chips from other companies, including Nvidia. The effort began about a year before Reuters’ July 2026 report, according to one of its sources. Meanwhile, the company is seeking to raise at least $12 billion in a second funding round that would increase the AI lab’s valuation by around 50% to RMB500 billion (about $75 billion). DeepSeek plans an IPO early next year, according to Bloomberg.