Summary
The release supports additional deployment formats and lets TensorRT engines serve a range of batch sizes through one optimisation profile. It also restructures export internals around a registry and removes several deprecated APIs.
Why it matters
Broader and more predictable export support reduces the engineering work required to move vision models from training into heterogeneous production runtimes.
Related coverage
- Claude Code 2.1.282 hardens telemetry, permissions and session recovery
- Payload CMS adds agent-assisted upgrades in its 4.0 canary
- Lightdash adds custom agent skills and exposes them through MCP