New endpoint
Step 2 of 2. Pick hardware and launch; your agents connect to the endpoint it gives you. Back to step 1
Model
Tuned recipes measured on real hardware · all runs
Models
Answers 200 when ready; every other path needs the endpoint key, checked at the edge.
Registry credentials are used once for the pull and never stored.
—
Weights as stored—
Parameters—
License—
Create a read token at huggingface.co. Used once to check access, kept for the build for 30 days, never logged.
Hardware
pick a model to check fitLaunch
Not readyThe same build on identical nodes under one address; requests are spread across them and the price multiplies. Sessions are not pinned to a node, so agents with long contexts gain more from one bigger node than from several small ones.
Change what it is tuned for
Normal machine memory, not GPU memory, used to keep your agents’ context cached. A larger cache keeps more agents’ context warm; leave it empty or 0 for none.
Advanced engine settings
Appended after Inferoptimal’s own flags; the model, served name, port and key stay managed.
1peak, not average128
Per-agent speed
Instances—
Agents per node at this speed—
Speed per agent—
Cost per 1M output tokens—
Context—
Measured
TTFT median
Per user
Total