A BIGAI-incubated team wants to write enterprise know-how into model weights

Tongyan Lingzhi (通衍灵智), a company incubated within the Beijing Institute for General Artificial Intelligence (BIGAI) ecosystem, is building what it calls Recursive Self-Improving Personal AI. QbitAI profiled the team on October 19, framing the bet simply: general capability will get cheap, and the differentiator will be whether an organisation's way of working can be folded into a model.

The architecture pairs a shared base model with personalised parameters. The base supplies general knowledge and reasoning, while individuals, projects or enterprises adapt it through lightweight Personal LoRAs that absorb validated professional preferences and behavioural patterns, rather than leaving them in prompts, retrieval stores or memory modules.

The substance of the approach is the learning loop. Task trajectories, human corrections and real outcomes become candidate training data; an Evaluator filters them against a specific user's or company's standards; the accepted samples drive incremental LoRA updates. Each updated model is then re-run against the previous version on the same tasks, shipping only if it clears the bar and rolling back if it regresses.

The Evaluator is itself part of the loop. Results from deployed updates feed back into how candidates are filtered and scored, creating a two-way iteration. That is why the company insists a single rejected answer does not trigger training: one piece of feedback is a signal, a recurring pattern is what becomes learnable.

The market argument is grounded in everyday friction. CEO Xie Xiaonan says enterprise demand is shifting from acquiring general capability to accumulating proprietary capability. Li Wei, CEO of BIGAI's industry platform, argues that stronger base models make personalisation more valuable, producing a division of labour between shared general intelligence and specialised, evolving models.

Cost is where it gets hard. If a company must maintain separate parameters for hundreds of roles and thousands of agents while supporting continuous training and live serving, conventional separate training and deployment becomes expensive. Tongyan Lingzhi therefore builds infrastructure too: TIP overlaps training with inference on the same GPUs using asynchronous gradient accumulation, while Paged LoRA manages memory so many adapters can share one base model.

Published interim tests give concrete numbers. On Qwen2.5-72B-Instruct with eight NVIDIA H800 GPUs, the team held the LoRA memory budget to 5.79GB in both paged and conventional flat modes. Paged mode showed advantages in latency and tail behaviour when LoRA requests were between 1 and 30 percent of traffic, but at 50 percent frequent parameter swapping still caused degradation.

The limits are stated plainly. CTO Chu Bingbing says the core RSI route and product architecture are defined and modules are being engineered, but a fully automated closed loop remains in development. Catastrophic forgetting, feedback contamination, model drift and update stability are all still open problems.