The official desktop build of DeepSeek Harness is now available for download on GitHub. The version number is 0.1.7-rc.1.20260924.1, corresponding to the official repository tag dsh-v0.1.7-rc.1. All core interfaces including the welcome page, local workspace, account balance panel, and about page are fully functional, not just placeholder mockups.

This desktop client uses your local folder directly as your workspace. You no longer need to juggle multiple browser tabs when researching data, creating spreadsheets, writing code, or troubleshooting. The official Desktop app is located under the apps/desktop directory in the repository. It builds on Electron and has Node, pnpm, and Python built in, so it can run even if you don't have these development environments pre-installed on your system.

Getting started takes just three steps: initial setup → select workspace → top up your credit balance. Tokens are deducted based on actual execution; you are not charged when no tasks are running. Although the installer hasn't been officially announced to the public yet, the community has already found the installable build on GitHub. Make sure to get it from the deepseek-ai/deepseek-harness repository, don't download packages from untrusted sources.




You can now finally run an Agent with just a double-click. Some netizens commented: "No more switching between browser tabs finally," while others said they were "already uninstalling my grok build right now." However, some users noted that the official client doesn't support a sidebar, which is a minor shortcoming.

On the same day the desktop client was released, Liang Wenfeng shared a set of figures during an investor meeting that explains another ongoing mystery: why high-end gaming graphics cards have been so hard to get.

The Information reported that DeepSeek's Annual Recurring Revenue (ARR) has surpassed $1 billion, just a few months after it was still under $500 million. This growth is mainly driven by recent API price hikes and the exploding popularity of DeepSeek's models. In terms of computing power allocation, over 70% is used for training new models, leaving less than 30% for inference.



Internal testing found that smaller models can handle most daily inference tasks perfectly well when run on consumer gaming GPUs. As a result, DeepSeek moved all its inference workloads onto gaming GPUs, so high-end data center training cards don't need to be pulled away from training to run inference. Market observers guess that these purchases specifically target models like the RTX 5090, which is why retail stock has remained extremely tight.


So it turns out that RTX 5090s aren't selling out to gamers—DeepSeek's inference workloads have soaked up all the supply of consumer gaming cards. Some netizens commented: "No wonder Liang has been buying up gaming GPUs," "All the good hardware is going to run models, that's why regular players can't get any." There are also skeptics who argue that 5090-class cards aren't powerful enough, and that DeepSeek must be using at least Pro 6000-level cards.

Running inference on consumer-grade graphics cards isn't a new trick, but this setup both eases DeepSeek's own computing shortage and coincidentally explains the supply-demand imbalance in the retail graphics card market. The desktop client brings Agent functionality closer to ordinary users, and the computing power allocation strategy shows how DeepSeek is keeping its operating costs under control. Looking at these two developments together, DeepSeek's strategy is pretty clear.