Breaking changes are labelled as such and announced ninety days ahead. We have never removed an endpoint without doing that and we do not intend to start.
You can now reserve a replica for a single workload in any region marked dedicated on the network page. A reserved replica is never shared, never cold, and billed monthly rather than…
Send prefix and suffix as separate fields instead of concatenating them into one prompt. The gateway assembles the sentinel tokens the model actually expects, which differ between releases and are easy to…
A proxy in front of three regions was buffering until 4 KB had accumulated, which added between 60 and 300 ms to the first visible token depending on how chatty the model…
Codes are now iad1, fra1, sin1 — an IATA code and an index — replacing the old us-east style. The old names keep working until 9 September 2026 and return a Deprecation…
A region field on the request overrides the deployment default for that call. Combined with residency: strict it will fail rather than route elsewhere, which is what you want when the data…
We were counting tokens against the total parameter count rather than the active experts, which overcharged Mixtral and DeepSeek V3 requests by roughly eight per cent. Affected accounts have been credited automatically;…
Weights are now memory-mapped from a local NVMe cache rather than pulled from object storage on first use. A model that has not been called in an hour now answers in about…
Per-region availability, p50 and p99, updated every thirty seconds, with an RSS feed of incidents. We publish the postmortem for anything that breaches the availability target, including the ones that were entirely…
One command, thirty-one regions, and an invoice that matches what you actually served.