Ex-Huawei Pangu lead Wang Yunhe founds RSI and unveils a 125B agent model

On November 5, Zhidongxi reported that Wang Yunhe, formerly the head of Huawei's Pangu large model effort, has founded a startup called RSI and released a 125-billion-parameter agent model.

The report gives two things: where a senior model lead went, and what the company says it is building. Wang Yunhe previously ran Pangu at Huawei and now appears as a founder, while RSI has written "agent model" directly into its product positioning.

A 125-billion-parameter agent model is a statement that agentic capability should live inside the model, not in a wrapper around someone else's API. If that number reflects the full weight set, the company is betting that base model capability sets the ceiling for what an agent can do, and is paying the training bill accordingly.

There is no settled definition of an "agent model." One approach puts an orchestration layer around a general model and lets the layer decompose tasks and call tools. Another writes multi-step planning, tool use and long-horizon execution into the training objective so the model performs those actions itself. The second is far more expensive to train and generally more stable on complex tasks.

Founders leaving core teams at large Chinese technology companies have become common, so the more interesting question is what they carry out with them. Training a large model requires data pipelines, distributed training engineering and evaluation systems as much as algorithmic insight, and that operational knowledge is harder to replicate than any single technique.

What the report does not say matters just as much: whether the model is open, its context length, its benchmark results, its training cost and the size of any funding round. Those details decide whether RSI's product is open weights developers can run themselves or a closed service reachable only through an API.

The agent space is already crowded, from coding assistants to enterprise workflow automation, and a 125-billion-parameter model cannot differentiate itself on size alone. The real dividing lines will be success rates on specific tasks, cost per completed task and how auditable the agent's actions are.

Three signals are worth watching: whether a technical report or model card follows, whether reproducible evaluations are published, and which industry the first named customers come from. Until then, 125 billion parameters is a directional claim rather than a verified capability.