v0.12.25 v0.28.0
vectorize-io/hindsightAug 31, 2026 → Aug 31, 20260 days between releases
What's New
Major vLLM 0.28 rebase adds AI models, speech, and diffusion support
Breaking Changes
- Rebased on vLLM 0.28, which may introduce breaking changes in serving behavior
New Features
- Scheduler-managed paged KV cache for diffusion models
- Full-duplex speech support for PersonaPlex and Nemotron
- Advanced MiniMax H3 deployment options (DLO, FP8)
- New TTS and Video generation models
Improvements
- Rebased on vLLM 0.28 for updated serving capabilities
- Expanded MiniMax H3 deployment options
- Expanded full-duplex speech serving
Upgrade Notes
Rebasing to vLLM 0.28 requires testing serving behavior. New diffusion and speech features may require updated configuration for model loading and resource allocation.
The Two Releases
v0.12.25
Aug 31, 2026· 2 months ago
A maintenance release closing a gap in version history, adding participant reading APIs, Teams passcode support, and resource sizing for pods.
Full release notesv0.28.0
Aug 31, 2026· 2 months ago
This release significantly expands model support and serving capabilities, introducing scheduler-managed paged KV cache for diffusion, full-duplex speech stacks for PersonaPlex and Nemotron, and advanced MiniMax H3 deployment options.
Full release notes