v2.2.18 v0.30.0
vectorize-io/hindsightSep 25, 2026 → Sep 25, 20260 days between releases
What's New
The newer release is a major vLLM-Omni update with a unified full-duplex framework and breaking configuration changes, unrelated to the NocoBase bug fixes in the older version.
Breaking Changes
- The `diffusion_batch_size` configuration parameter has been removed; use `max_num_seqs` instead.
New Features
- Unified full-duplex serving framework with engine-owned sessions
- MAGI-2 Preview model for text/image-to-video-and-audio
- Tencent AuK and AuK-Flash for speech generation
- Opt-in multi-process API serving via `--api-server-count`
Improvements
- Native cross-stage KV and multimodal payload transfer
- Interactive world-model and streaming video generation capabilities
- Expanded Ascend NPU support with Wan2.2 MXFP4
Upgrade Notes
The `diffusion_batch_size` configuration parameter has been removed and must be replaced with `max_num_seqs`. The release introduces a unified full-duplex serving framework and expanded hardware support.
The Two Releases
v2.2.18
Sep 25, 2026· 27 days ago
A bug fix release for NocoBase addressing issues with workflow forms, approval sub-tables, and Grid Card pagination.
Full release notesv0.30.0
Sep 25, 2026· 26 days ago
A major release for vLLM-Omni featuring a unified full-duplex serving framework, new model support (MAGI-2, MiniMax H3), and expanded hardware support.
Full release notes