v2.2.0 v0.28.0
HKUDS/Vibe-TradingAug 24, 2026 → Aug 26, 20262 days between releases
What's New
Switched from Spice.ai v2.2.0 to VLLM v0.28.0, gaining Kimi-K3 optimizations and DeepSeek V4 support.
Breaking Changes
- bitsandbytes support migrated to an out-of-tree plugin
- Transformers bumped to 5.15.0
- Removed deprecated `calculate_kv_scales` runtime KV scale calculation
- Removed `override_attention_dtype`
New Features
- Kimi-K3 performance push with DCPU and FlashKDA
- DeepSeek V4 sparse MLA support and AMD Quark NVFP4 support
- Model Runner V2 disaggregation and weight offloading
- Tiered KV cache offloading support
- Rust frontend and gRPC support
Improvements
- Significant performance push for Kimi-K3
- Speculative decoding advances with DFlash2 and DSpark
- ROCm enablement on gfx11 and gfx950
Upgrade Notes
Update bitsandbytes to the new plugin, upgrade Transformers to 5.15.0, and remove deprecated runtime KV scale calculation and attention dtype override configurations.
The Two Releases
Spice v2.2.0 (August 24, 2026)
Aug 24, 2026· 2 months ago
A major release focusing on real-time data capabilities and performance. It introduces Cloud Connect for self-hosted runtimes, MySQL and PostgreSQL Change Data Capture (CDC) support, and warm in-memory indexes to accelerate search.
Full release notesv0.28.0
Aug 26, 2026· 2 months ago
Major performance and feature release for VLLM including Kimi-K3 optimizations, DeepSeek V4 support, and KV cache offloading.
Full release notes