The news
DeepSeek, the Chinese AI developer, released version 0.2 of its DeepSeek Harness agent on September 29, 2026, with its first desktop installers for macOS and Windows, according to a write-up by NYU Shanghai's Research IT team (RITS). The DeepSeek Harness desktop app lets a user sign in, pick a local folder as a workspace and hand an AI agent a task, without setting up developer tools first.
DeepSeek's product page describes Harness as open source under the MIT license and in public preview, aimed at everyday tasks, coding, research, background automation and plugins that users can install or create through chat. Its default model is DeepSeek-V4.1-Flash. Users sign in with a DeepSeek account or API key, and the app also supports other providers including Anthropic, OpenAI, Kimi and GLM, according to RITS.
The risks are spelled out. RITS reported that the agent runs shell commands, loads third-party plugins and can modify or delete files and disclose data or credentials within its workspace. DeepSeek says Harness has not undergone a security audit and should not be treated as secure or production-ready, and recommends least-privilege setups, disposable virtual machines or containers, and backups, RITS reported. The agent's reasoning is processed in DeepSeek's cloud by default; only voice transcription runs locally.
Chinese models are also reaching desktops without any cloud at all. Strata, an open-source inference engine on GitHub, says it runs Qwen3.8-Flash-Next, a model from Alibaba's Qwen family, on consumer gaming PCs with an Nvidia RTX or AMD Radeon GPU. Its model guide says system RAM must cover the model's expert weights plus about 10 GB, so the smallest version needs roughly 44 GB and 48 GB fits the two smallest, with downloads of 66 to 76 GB. The project claims 94 tokens per second on a 12 GB Nvidia RTX 5070 paired with 64 GB of RAM. Strata is a community project, not an Alibaba release.
Developers are testing these models for real work. In an October 2, 2026 blog post, Thibaud Colas of the Wagtail open-source content management project described a month of trying to code using only GLM-5.3 Flash, an efficient open model available from multiple providers. The team used 2 billion tokens in total but only half went to the target model, and it reported $68 of GLM-5.3 Flash usage, within its budget.
The numbers
- Harness v0.2 release
- September 29, 2026
- GPU memory on tested RTX 5070 (Strata)
- 12 GB
- Approx. minimum system RAM, smallest version (Strata)
- About 44 GB
- Claimed speed on RTX 5070 (Strata)
- 94 tokens per second
- Wagtail tokens used / share on GLM-5.3 Flash
- 2 billion / 50%
Why CEOs should care
For CIOs and CISOs, the problem is visibility. A free, open-source desktop agent does not arrive through a purchase order, so vendor reviews and data processing agreements never get triggered. Yet DeepSeek's own guidance says Harness can touch files and credentials in its workspace and sends reasoning to DeepSeek's cloud by default. Ask: can we see which AI agents are installed on company devices, and which endpoints they call?
Locally run models create the opposite blind spot. A Qwen model running on a gaming GPU sends no traffic to an AI provider, so network monitoring will not flag it, but it can still process confidential documents and generate code that enters production without review. Acceptable-use policies should cover local models explicitly, not just cloud chatbots, and endpoint controls should govern who can install model runtimes.
Boards and general counsel should also weigh origin. Companies in regulated or defense-related sectors may face customer or contract terms on where data goes. The practical step is a clear list of approved models and tools, with a fast review path so employees who want cheaper, capable models have a sanctioned option.
The bigger picture
Cheap, efficient models are making AI a commodity that individual employees can adopt on their own, much as cloud apps did a decade ago. The Wagtail experiment shows developers actively testing open models for cost and energy savings. Governance programs built around a handful of enterprise vendors will need to cover this long tail.
What’s next
Harness is still a developer preview. Watch for a security audit, enterprise controls such as managed deployment and logging, and whether endpoint security vendors add detection for local model runtimes like Strata.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story








