v1.23.0 v0.30.0
HKUDS/Vibe-TradingSep 22, 2026 → Sep 22, 20260 days between releases
What's New
Switch from PDF text repair to LLM inference engine with new models and optimizations
Breaking Changes
- `--enable-scale-out` is now opt-in on plain `vllm serve`
- `g_idx` activation ordering removed
- Deprecated items removed, including env vars and cache modes
- `all` Mamba cache mode deprecated
- `python -m vllm.entrypoints.grpc_server` deprecated
New Features
- DeepSeek-V4.1-Flash and GLM-5.3-Flash models
- Fast Start IPC cache daemon for persistent GPU weight caching
- HiSparse host-resident tier for sparse-MLA decode
- Gumbel-max watermarking with keyed PRF
- K2-Horizon and Cohere Compass model support
Improvements
- Model Runner V2 dual-batch overlap and FULL CUDA graphs
- Targeted online quantization and W4A16 DSA support
- Improved performance for Qwen3.8-Flash-Next
- Async Engram prefetch with DP sharding
- Rust frontend forwarding per-request controls
Upgrade Notes
Enable `--enable-scale-out` explicitly if scaling is required. Use `vllm serve --grpc` instead of the deprecated module. Remove deprecated environment variables like `VLLM_PREFIX_CACHE_RETENTION_INTERVAL`.
The Two Releases
1.23.0 — Text-layer repairs
Sep 22, 2026· 1 month ago
This release focuses on repairing text layers by handling malformed bounding boxes, improving right-to-left text reading order, and fixing spacing for accents and detached glyphs.
Full release notesv0.30.0
Sep 22, 2026· 1 month ago
A major update featuring 762 commits introducing new models like DeepSeek-V4.1-Flash and GLM-5.3-Flash, significant performance optimizations for Model Runner V2 and Qwen3.8-Flash-Next, and new features like Fast Start IPC caching and HiSparse host-resident sparse decoding.
Full release notes