Size a setup
You give the price builder a head count, the kind of work and the models you want. It works out the nodes.
Work has a weight
A stream is one conversation a model answers at 26 tokens a second or more. Each kind of work needs a share of a stream per person at the busiest moment of the day: the person's own conversation, plus agent work that runs without them.
| Work | Example | Streams a person, at peak |
|---|---|---|
| Documents and research | Reports, contracts, email | 0.46 |
| Software | Coding agents on real repos | 0.67 |
| Agents | One running agent per person | 0.67 |
| Support and questions | Many short answers | 0.25 |
A team that does several kinds of work switches between them, so a person's own conversation counts once and the agent work is averaged. Agents are held to one running agent per person.
From streams to nodes
- Streams. People times their weight, rounded up, shared between the models you pick.
- Copies. Each model gets enough copies for its share. A model too big for one node spreads each copy across linked nodes. Small models share a node when memory and speed allow.
- A range. Every setup is worked out twice, with what we measured and with a careful margin, so the price shows as a low and a high.
The price page covers 15 to 100 people on up to 12 nodes. Every price it shows saves you at least 10% against $100 seats over three years. A bigger setup, Kimi K3.1, or a setup that needs more hardware than that saving allows, is priced on a call.