New
B200 capacity generally available in nine regions, for both serving and multi-node training. Training reservations include provisioned interconnect.
Private Marketplace listings. Publish an internally-built agent, connector or evaluation suite to your own organisation with versioning, access control and one-click deployment.
Per-case promotion gates. An evaluation suite can now designate individual cases as hard gates — failing one blocks promotion regardless of the aggregate score. Previously only the aggregate threshold could block.
Distillation pipelines. Capture production frontier-model traffic with human corrections and tune a small model on it, as a managed pipeline rather than an export-and-reimport exercise.
Improved
- Entity resolution throughput improved roughly 2× on corpora above 50 million records.
- Survivorship rules can now be expressed per attribute per jurisdiction.
- Hybrid retrieval fusion weights are tunable per corpus rather than per estate.
- Agent tool registers now surface unused tools, which is a useful prompt to remove surface area.
Fixed
- Fixed quarantine records being re-evaluated on every ingest cycle rather than on change.
- Fixed a case where the abstention threshold was not applied to hybrid retrieval results.
- Corrected timezone handling in scheduled evaluation runs for estates spanning regions.
Deprecated
| Item | Replacement | Removal |
|---|---|---|
| Aggregate-only evaluation thresholds | Per-case hard gates | 2026.12 |
| Estate-level fusion weights | Per-corpus weights | 2026.10 |
- Release
- Compute
- Marketplace
- Evaluation