This time, it’s for real: after several delays, Oracle AI Database is finally available on standard hardware. For now, on x86-64 servers. This concerns the Enterprise and Free editions.
In parallel, branding is evolving: goodbye to Oracle AI Database 23ai, hello to version 26ai. Between one and the other, the internal architecture does not change. The APIs do not either. There is therefore no need for an upgrade, nor for recertifying the applications. The LTS status is preserved and with it, the support policy (the end of the first phase on December 31, 2031).
Migration from versions 19c and 21c can be performed without going through 23ai.
Vectors, indexes, algorithms… Oracle beefs up its vector search
This marks the first time that an “AI-branded” release is available on-prem, outside Oracle systems (Exadata, ODA, PCA). Although some features of 23ai have been backported to 19c.
Among the new features in the 26ai release:
- Binary vectors and sparse vectors
- New vector distance metric (Jaccard)
- Disk checkpoints to accelerate the reconstruction of in-memory HNSW indexes
- Automatic reorganization of IVF indexes
- Management of ONNX models as first-class objects
On the security front, the SQL firewall — which requires a dedicated license — is now included with Oracle Database. The 26ai release adds TLS 1.3 support and simplifies the protocol deployment (clients no longer necessarily need to provide a root certificate store, notably). TDE encryption moves to AES-256 by default, and the maximum password length increases from 30 to 1024 bytes. Post-quantum cryptography arrives, with ML-DSA for certificate signing and ML-KEM for key exchange (potentially hybridized with ECDHE).
RAC (Real Application Clusters) becomes deployable in container environments. Meanwhile, patching is split into two phases (preparation, activation) to minimize the impact on availability.
