Something remarkable has happened in artificial intelligence over the past eighteen months, and it deserves to be said without hedging: Chinese AI labs are producing some of the finest engineering in the world, and they are giving it away. GLM 5.2, released this June under an unrestricted MIT license, is the strongest open source model ever published, ranking top three on major coding leaderboards alongside the best closed American systems. The Kimi K2 line, the DeepSeek V4 family, and MiniMax M3 fill out an open source top tier that, a year ago, did not exist at anything close to this level.
For American users and businesses, this is genuinely good news with one genuinely serious asterisk, and the asterisk is almost never explained clearly. It has nothing to do with the models themselves, and everything to do with where you use them. Phoenix Grove AI, (found at www.pgsgrove.com) is making it not only safe to use them, but powerful.
What Open Weights Do and Don’t Protect
An open weights model is a published artifact. When a lab releases weights under an MIT license, anyone can download, inspect, modify, and run them. The model cannot spy on anyone from a hard drive. It cannot phone home. It is, in the most literal sense, just a very large file of numbers, and the moment it is published, no government or company can take it back. Weights released this June already exist on storage across every continent. That permanence is precisely why open models became so attractive the same month American frontier access proved conditional, with one leading closed model pulled from the market days after launch under an export directive and another lab’s flagship released only to a short list of approved organizations.
But here is what the license does not cover: your conversations. A frontier-scale model has hundreds of billions of parameters and does not run on consumer hardware. Practically speaking, using GLM 5.2 or its peers means using someone’s hosted service, and your data lives under the laws of wherever that service operates.
When the host is the developer’s own app or API, a user’s conversations live under the laws of the country where those servers sit. For many American users and businesses, that alone is the whole issue, and it requires no claims about anyone’s intentions. Data residency is a standard requirement in cloud computing for precisely this reason: people and organizations generally want their information governed by their own legal system, with the recourse, transparency norms, and privacy expectations they know. A chat history stored abroad answers to foreign law, whatever any privacy policy says, and for anyone doing real work, that question deserves a deliberate answer rather than a default one.
The mirror-image discomfort exists on the American side, where the concern is corporate rather than governmental. Big AI consumer services increasingly tie useful features like memory to training consent, retain reviewed conversations for years, and have begun introducing advertising into answers. Neither direction of the squeeze is invented, and the fact that both are real at once explains why so many users describe feeling cornered.
Separating the Model From the Serving
The way out of the corner rests on one idea that most coverage of open source AI never quite states: the model and the serving are separable. Because the weights are open, anyone with sufficient hardware, anywhere in the world, can serve these models under their own jurisdiction and their own data practices. A Chinese-developed model served from American infrastructure under American law, by a company with no training pipeline, offers a fundamentally different privacy proposition than the same model served from its developer’s servers, and an arguably better one than closed American services that harvest by default.
That separation is the entire premise of the Open Grove, the open source wing of PGS AI (ai.pgsgrove.com), a platform built by the independent American company Phoenix Grove Systems. The Open Grove serves GLM 5.2, the Kimi K2 line, the DeepSeek V4 family, MiniMax M3, and more, and according to the company, every conversation is processed on privacy-first US-based infrastructure with zero data retention at the inference level. User prompts never route through the original model developers’ servers at any point. Conversations are never used for training, a capability the company says it deliberately never built, and no conversation data is sold or shared.
“You can admire the engineering and still want your data on American soil. Those aren’t contradictions,” the company’s founder said. “The whole gift of open source is that you don’t have to choose. The model comes to you. Your life doesn’t go to it.”
Serving Openly, Serving Carefully
Phoenix Grove’s approach adds two layers that address the subtler concerns about imported models. Every model in the Open Grove runs with the company’s ethical grounding and bias mitigation layer, a calibration approach developed in-house to help models, whatever their origin, respond to human diversity with dignity. And the platform displays each open model’s reasoning traces in full, so users can watch how a model actually worked through their request. For anyone harboring vague unease about a foreign model’s behavior, visible reasoning is a more useful answer than reassurance.
Around the models, the platform provides what the official apps famously lack: persistent long-term memory across every conversation, a voice interface in which the user’s audio is transcribed locally and never leaves their device, collaborative canvases, web research, and isolated workspaces. A single subscription, starting at four dollars per month with the first month free, covers both the Open Grove and the company’s own proprietary cognitive builds. Developers who want raw access can use the company’s separate API at api.pgsgrove.com, which carries the same US processing and zero retention posture.
The Effect For Ordinary Users
The practical guidance that falls out of all this is short. If you are experimenting casually with an open model and sharing nothing sensitive, the official apps are a fine way to look. If you are doing real work, personal, professional, or anything in between, treat the hosting question as seriously as the capability question, because your data’s legal home is determined by the former and not the latter. Ask any AI service, American or otherwise, three things: what jurisdiction holds my conversations, are they retained, and are they trained on.
The deeper takeaway is more hopeful than the geopolitics suggests. The open source surge means the best AI engineering now belongs to everyone, including the ability to serve it safely. Chinese labs built remarkable models and published them freely. American infrastructure can serve them privately. Users, for the first time, can have both halves of that sentence at once, and it took a small independent company to connect them.
