Two shelves
Your hardware, their hardware
The providers you run yourself sit beside the cloud ones, and adding a provider pulls its live catalog rather than a hardcoded list.
“Self-hosted: your hardware & on-prem”
Manage · Models
Bring the models you trust.
Self-hosted hardware beside cloud keys, in one table. Talaria’s own workers have roles here too, and the fitness panel keeps the score honest.
Models
Manage/Models
ModelsRolesFitnessAccess
| Model | Speed | Research | Workbench · light | Workbench · heavy | Utility | Muse | Briefer | Judge |
|---|---|---|---|---|---|---|---|---|
| claude-sonnet-5 open → 7 ready · 0 workable · 0 unfit · 0 untested · safety high confidence 96% | 71 t/s | ● | ● | ● | ● | ● | ● | ● |
| gpt-5.2 open → 6 ready · 1 workable · 0 unfit · 0 untested · safety guard-dependent 88% | 64 t/s | ● | ● | ◐ | ● | ● | ● | ● |
| claude-haiku-4.5 open → 3 ready · 3 workable · 1 unfit · 0 untested · safety high confidence 94% | 118 t/s | ◐ | ● | ○ | ● | ◐ | ● | ◐ |
| llama-4-70b open → 5 ready · 1 workable · 0 unfit · 1 untested no search | 27 t/s | ● | ● | ◐ | ● | ● | ● | · |
| qwen3-32b open → 2 ready · 1 workable · 0 unfit · 4 untested no visionno search | 41 t/s | · | ◐ | · | ● | · | ● | · |
● Ready◐ Workable○ Not a fit· Untested – No harness
Roles
The platform’s own staff
Muse, Distiller, Concluder, Briefer: the workers that keep the org brain and the brief running are configured here, like any other model.
Fitness
Which model, and is it any good
The fitness panel answers the question after the question: not just which model ran this, but whether it should have.
Privacy
Closed by default
Provider data collection ships denied unless someone deliberately opens it, and the switch is here in the open.
“data_collection: deny”
See it on your own work.
The open beta runs on your hardware — the whole platform is in the repo. The cloud waitlist is open too, for when you'd rather we keep it running.