Building agents that take real actions means picking a model you can trust, afford, and control. After working through the options we chose Qwen’s QwQ-32B and ran it on our own infrastructure. Here is the reasoning — the trade-offs, not the hype.
Why QwQ-32B
- Strong reasoning — the model is built for step-by-step problem solving, which is what agent planning needs.
- Open weights — we can inspect, fine-tune and deploy it without asking permission.
- Right-sized — capable enough for the work, small enough to serve economically.
Why self-host instead of an API
- Data control — sensitive inputs never leave our environment.
- Predictable cost — at agent volumes, per-token API pricing adds up fast; owned infrastructure flattens it.
- Latency — co-locating the model with the workload cuts round-trips.
- No lock-in — our stack doesn’t depend on one vendor’s roadmap or rate limits.
The trade-off we accepted
Self-hosting isn’t free — you own the GPUs, the serving stack and the on-call. We took that on deliberately, because for production agents handling real data, control and cost predictability mattered more than the convenience of a managed endpoint.
“AI without giving up control of your data. Your cloud, your models, no lock-in.”
How it’s built
.png)
.png)
