Clients often ask which AI model we use. Our usual answer: it depends.
Why we don't commit to one
Different models are good at different things. One handles long documents well, another writes more natural Chinese, another is cheap and fast enough for large volumes of simple work. Tying a project to a single provider means giving up those choices before you start. Models also move quickly: today's best may not be the best in a few months.
How we choose
For a given task, we test several models against the same set of real examples, then compare four things: quality, cost, speed and compliance. For projects in China we often use DeepSeek, Qwen, Kimi, Zhipu GLM and Doubao; for projects elsewhere, OpenAI, Claude, Gemini and Llama. It's common for two or three models to share the work within one system.
How we handle data
- We prefer enterprise APIs that don't train on customer data
- Where data is sensitive, we deploy open models on the client's private servers or their own hardware
- We connect to data with only the permissions needed, and the contract spells out how it's used
- Data that can be processed locally isn't uploaded
Leaving room for later
We design systems so that switching models is cheap. When a better model comes along, we plug it in and run the same examples, and if it holds up, it goes into use.