Open-Weight Models
Your models, your machine, your audio.
Dograh is the orchestrator. You choose what runs underneath, and open weights on hardware you own are the only choice that keeps the call inside.
01 / THE LAYER THAT MATTERS
A voice agent has two layers. Orchestration holds the workflow, the turn-taking, the tool calls and the telephony. Inference is speech to text, the model that decides what to say, and the voice that says it. Dograh is the orchestrator. It is not a model, and it does not change when you change what runs beneath it.
02 / WHERE THE AUDIO GOES
Self-hosted open weights against a hosted platform
03 / WHY IT HOLDS UP
The objections that used to stop this
Latency is answered
Putting every stage on one machine removes the network hop between them rather than shortening it.
Sovereignty by architecture
Bring your own key moves the commercial relationship. The audio still reaches the provider.
One ordinary GPU
A 4-bit 27B language model and a small speech model fit together on a single 24GB card.
Swap the layer underneath
Moving inference onto your own hardware is a configuration change, because the orchestration layer does not move with it.
04 / MEASURED, NOT ARGUED
Questions
Does bring your own key give me data sovereignty?+
No. It changes who holds the billing relationship. The audio still travels to the provider and is processed on their infrastructure.
Is Dograh a voice model?+
No. Dograh is an orchestrator: workflow, turn-taking, tool calls, telephony and call lifecycle. Inference is a separate layer you choose.
What hardware do I need to run open weights?+
A single 24GB GPU is enough for a 4-bit 27B language model alongside a small speech-to-text model.
Will self-hosting make my agent slower?+
No. Colocating every stage on one machine removes the network hop between them rather than shortening it.
Can I start hosted and move to open weights later?+
Yes. The orchestration layer does not change, so moving inference onto your own hardware is a configuration change.
Are open-weight models the same as open source?+
No. Open weights means you can download and run them. Licences vary, and some forbid commercial use, so read each one.
Keep the orchestrator. Change the models later.
Start on hosted inference to get a workflow live, then move it onto your own hardware when compliance catches up with the pilot.