v2026.824.1 v0.28.0
karpathy/nn-zero-to-heroAug 25, 2026 → Aug 26, 20261 day between releases
What's New
vLLM v0.28.0 adds Kimi-K3 optimizations and ROCm support; Paperclip fixes background service onboarding.
Breaking Changes
- bitsandbytes support migrated to an out-of-tree plugin
- Transformers bumped to version 5.15.0
- The deprecated `calculate_kv_scales` runtime KV scale calculation was removed
- The `override_attention_dtype` configuration was removed
New Features
- DFlash2 speculative decoding with local convolution
- Default prefix caching enabled for Mamba models
- Standalone renderer and multimodal image inference over gRPC
- Tiered KV cache offloading support with disk offloading
Improvements
- Kimi-K3 performance push with Decode Context Parallel (DCP) and fused kernels
- DeepSeek V4 sparse MLA support and AMD Quark NVFP4 support
- Model Runner V2 expands to encoder-only workloads and attention-free models
- Improved diagnostics for `paperclipai doctor` to identify missing service binaries
Bug Fixes
- Background service now materializes the managed install before registration
- Onboarding flow skips foreground start prompts if the service is already running
- Onboarding now ends by presenting the dashboard URL in the browser
Upgrade Notes
vLLM users must update Transformers to 5.15.0 and migrate bitsandbytes to the plugin system. Check configuration files for the removed `override_attention_dtype` and `calculate_kv_scales` flags.
The Two Releases
v2026.824.1
Aug 25, 2026· 2 months ago
A patch release for Paperclip that repairs the background-service leg of onboarding to ensure the service actually starts and the dashboard is correctly presented.
Full release notesv0.28.0
Aug 26, 2026· 2 months ago
A major performance release for vLLM featuring optimizations for Kimi-K3 and DeepSeek V4, Model Runner V2 maturation, and expanded hardware support including ROCm.
Full release notes