v0.28.0 v0.12.26
vectorize-io/hindsightAug 31, 2026 → Aug 31, 20260 days between releases
What's New
Newer release adds Zoom integration improvements and text-only STT, while older release focused on serving engine upgrades and speech stacks.
Breaking Changes
- Rebase onto vLLM 0.28 in older release (potential compatibility issues)
- Text-only STT responses may alter transcript data formats in newer release
New Features
- Per-participant speaker attribution in Zoom transcripts
- Improved Zoom admission via browser-signal hardening
- Dismissal of in-meeting popups during calls
- Text-only STT responses reaching transcripts
- Voxtral/JSON-only transcription backends
Improvements
- Speaker attribution replaces mixed-lane guessing
- Browser-signal hardening for Zoom admission
- Popups no longer obstruct recording
- STT backend flexibility with Voxtral/JSON-only options
- Detailed join failure logging with typed reasons
Bug Fixes
- Deaf-capture guard restored on per-track branch
- Mic advisory, AI-Companion notices, and hardware toast dismissed
- Name resolver fallback logic and self-exclusion backstop implemented
Upgrade Notes
For the newer v0.12.26 release, no specific configuration changes are required, but be aware that Zoom transcripts will now be named per participant. For the older v0.28.0 release, users should verify compatibility with vLLM 0.28 following the rebase.
The Two Releases
v0.28.0
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 notesv0.12.26
Aug 31, 2026· 2 months ago
Improves Zoom integration with better speaker attribution and admission rates, and adds support for text-only STT responses.
Full release notes