v2.2.1 1.10.0
Textualize/textualSep 2, 2026 → Sep 4, 20262 days between releases
What's New
Major version jump: Training speed improved 44%, but defaults and validation logic have changed.
Breaking Changes
- Validation evaluates one model per epoch instead of two
- grad_accum_steps defaults to 1 instead of 4
- Optimizer parameter groups collapse from one-per-parameter to one-per-hyperparameter
- Dataset builders now validate their config instead of silently falling back
- log_per_class_metrics defaults to False
New Features
- GPU batched linear-assignment solver for the matcher
- Optimizer parameter group merging
- Dataset builder validation
Improvements
- ~44% faster training time
- Up to 46% faster inference on large input frames
- 1.25–2.26× DataLoader throughput
Bug Fixes
- Query federation and SQL issues
- Validation evaluates one model per epoch instead of two
Upgrade Notes
Validation now evaluates one model per epoch, so metrics reflect the EMA model. Default grad_accum_steps changed from 4 to 1, effectively halving the batch size. Dataset builders now validate config instead of silently falling back.
The Two Releases
Spice v2.2.1 (September 1, 2026)
Sep 2, 2026· 2 months ago
A patch release improving Cayenne storage management, applying retention policies to more datasets, and fixing query federation and SQL issues.
Full release notesv1.10.0: Faster Training & Inference
Sep 4, 2026· 2 months ago
A training-focused release significantly improving speed with a 44% reduction in training time and introducing a GPU batched matcher.
Full release notes