Tongyan Lingzhi moves AI personalization from prompts into the model layer itself

Chinese startup 通衍灵智 (Tongyan Lingzhi) says AI should be personalized not with prompts or an external knowledge base, but at the level of the model itself. The company's slogan: from "one model for ten thousand people" to "a thousand faces for a thousand people." The available excerpt of the Jiqizhixin article gives no figures, model sizes or product names. So we look at how the problem is framed and at what to check once such claims reach real pilots.

A smart model with no memory of its own

The article's diagnosis is a familiar one. LLM reasoning keeps improving, context windows keep getting longer, and agents are moving into code development, data analysis and company production processes. But once AI shifts from tool to executor, a question arises: how does a single, ever-smarter general-purpose engine develop long-term capabilities that belong to a specific person, project or organization?

The authors give the example of a corporate agent. It can find any regulation and any piece of business knowledge. The available text cuts off after this, but the logic is clear: access to information is not the same as absorbing it. The agent knows what the documents say, but not how this team usually settles disputed cases, why it abandoned the previous architecture, or whose edits tend to pass review. That is my inference from the framing, not a quotation.

What "personalization at the model layer" means

Today, individual customization is usually assembled from several layers:

• system instructions and prompts;
• document search (RAG);
• agent memory and long context;
• fine-tuning, for example LoRA adapters.

The phrase "model layer" in the headline points to the last item: the user's experience should end up in parameters or in adaptive components, not sit outside as text notes. How exactly this works at 通衍灵智 cannot be determined from the excerpt. It could be adapters, continual learning or something proprietary.

This approach comes with well-known engineering difficulties:

• forgetting previously learned material when fine-tuning on new data;
• the cost of storing and serving thousands of individual versions;
• privacy: a person's data ends up inside the weights;
• quality evaluation when there is no reference "correct answer" for a specific team.

If the company solves even some of these, the thesis is interesting. If not, it becomes yet another nice slogan about "personal AI."

What the publication tells us

• Company: 通衍灵智, China.
• Thesis: personalization should happen in the model, not only around it.
• Scale of personalization: a person, a project, an organization.
• Context: agents in coding, analytics and corporate processes.
• Not disclosed in the available excerpt: architecture, benchmarks, prices, customers, timelines.

How to evaluate such claims

Anyone choosing a platform for corporate agents should ask any vendor making a similar promise for specifics. What exactly gets updated: the weights, an adapter or a memory layer? Can an individual version be rolled back? What happens to the data when an employee leaves or a contract is terminated? How is the gain measured against a "base model plus RAG" setup?

The last question is the key one. Retrieval-based setups already cover most tasks that require access to corporate knowledge. Personalization in the model is justified where style, habitual decisions and accumulated working context matter, and these do not fit well into documents. Until 通衍灵智 shows comparative tests, this is a hypothesis, even if a well-posed one. It is worth following the company's publications and independent evaluations precisely because of this distinction.