Serving
Scale to zero or always warm: choosing a serving mode
Should my model endpoint scale to zero or stay always warm?
Keep it warm if a human is waiting for the response. Let it scale to zero if a batch job, an internal tool or an automation is calling it. The decision is about who experiences the cold start, not about the size of the saving.
What each mode actually costs
An always-warm endpoint bills continuously for the time it is resident, regardless of traffic. That is the honest description of reserved capacity, and its virtue is that the bill does not change when usage does.
A scale-to-zero endpoint costs nothing while idle and pays a cold start on the first request after a quiet period. The saving is real and can be large for intermittent workloads.
The decision rule
Ask who waits. If a person is sitting in front of an interface, keep it warm — a cold start reads as the product being broken, however good the steady-state latency is. If the caller is a nightly job, an automation or an internal tool, let it scale to zero.
The saving is not the deciding factor. A cheap endpoint that makes your product feel broken is not a saving.
Setting cooldown properly
Cooldown is how long the endpoint stays resident after the last request, and it is the dial that decides how often cold starts happen at all.
Set it below the typical gap between requests and normal pauses trigger reloads — the worst outcome, because you pay for the reload and the user waits for it. Set it far above the quiet period and you have built an always-warm endpoint with extra steps.
Look at the actual distribution of gaps in your traffic. Set cooldown above the typical gap and below the quiet period.
Bound the queue explicitly
A cold endpoint queues requests while capacity is acquired. Both bounds are yours: a depth beyond which requests are refused with a retry hint, and a wait beyond which they time out.
Setting these is not pessimism. An unbounded queue turns a capacity problem into a mysterious latency problem, and the client never learns which it is facing.
Streaming hides a lot
With streaming enabled, the perceived latency is the time to the first token. That does not remove a cold start, but it does mean that once capacity is up, the response begins immediately rather than after the full generation.
For borderline cases — an internal tool that people do sometimes wait on — streaming plus a generous cooldown is often enough to make scale-to-zero acceptable.
Frequently asked questions
Can I warm an endpoint before a known burst?
Yes. Warming ahead of a scheduled job gets the best of both modes: no cold start when the burst arrives, no charge for the hours before it.
Does scale to zero affect adapters?
Adapters load against an already-resident base much faster than a whole model loads, so an endpoint serving several adapters recovers more cheaply than one serving several merged models.
Terms used here
More on serving
Last verified 6 August 2026.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.