Models of your choice
Any model you choose. Switch without rebuilding.
Open-weight models on your hardware, small models fine-tuned on your domain, or a hosted provider where you prefer. Every call goes through one gateway, so changing models is configuration.
- Open-weight
- Qwen, Mistral and others
- Fine-tuned
- Small models, weights yours
- Hosted
- Only if you opt in
- Gateway
- LiteLLM, cost per department
In short
DeepVox is model-agnostic: it runs open-weight models such as Qwen and Mistral on your own hardware, small models fine-tuned on your domain, or hosted model providers if you choose to use them. Every model call passes through a gateway (LiteLLM), so switching models is a configuration change rather than a rebuild, and usage and cost are tracked per department.
public/videos/platform/models.mp4 Then run python scripts/video-posters.py for the cover image. Three kinds of model
Pick per task, not per vendor.
Open-weight models
Qwen, Mistral and other open-weight families, served on your infrastructure or in your DeepVox cloud environment.
Fine-tuned small models
Compact models trained on your domain where they beat retrieval alone. Faster, cheaper to run, weights yours.
Fine-tuning →Hosted providers
A hosted model for a specific task, only if you switch it on, routed and logged through the same gateway.
The model gateway
One gateway, no lock-in.
Every call routed
Assistants, voice agents and automations call models through one gateway, never directly.
Rules per task
Choose which model handles which task, and which teams may use hosted providers at all.
Cost and usage visible
See model usage and cost per department.
Swap by configuration
A better or cheaper model arrives; you switch it in the gateway without rebuilding anything.
How we choose
Measured on your questions.
| Need | Usually the best fit |
|---|---|
| Data must never leave your environment | Open-weight model on your infrastructure |
| Narrow, repeated task or strict latency (voice) | Fine-tuned small model |
| Answers from changing documents with citations | Retrieval with an open-weight model |
| A specialist task where a hosted model clearly wins | Hosted provider, opted in for that task only |
Candidate models are benchmarked against an evaluation set built from your own questions, and that set stays with you.
FAQ
Questions, answered.
Which models does DeepVox use?
Our own models by default, built on open-weight families such as Qwen and Mistral, chosen in the mapping step against your task and hardware. Hosted providers are only used if you opt in.
Can we bring our own model?
Yes. If a model can be served on your infrastructure or reached through the gateway, it can be used and measured alongside the others.
What happens when a better model comes out?
It is evaluated on your test set and, if it wins, switched in through the gateway. No rebuild.
Are our prompts used to train someone else’s model?
Not with open-weight or fine-tuned models on your infrastructure. For any hosted provider you opt into, we flag its data terms before you switch it on.
Related
Fine-tuned small models
Your domain, your weights.
Learn more →RAG vs fine-tuning
How we decide on evidence.
Learn more →Open-source stack
Serving, gateway and evaluation tools.
Learn more →Start with two weeks of evidence, not a sales call.
A fixed-price discovery sprint, credited against whatever comes next. Or write to sales@deepvox.ai.