v0.55.0
Mesh-LLM/mesh-llmv0.55.0Apr 3, 2026by github-actions[bot]
AI Summary
Focuses on asymmetric KV caching for larger models, runtime model management, and critical connection stability fixes.
Key Highlights
- Implemented asymmetric KV cache (Q8_0 K + Q4_0 V) for 5-50GB models
- Added runtime model load/unload via CLI and REST
- Fixed critical client --auto join hang (connect_to_peer regression)
- Added context-aware host selection and retry logic
New Features
- Asymmetric KV cache
- Runtime model load/unload
- Context-aware routing with retry logic
Full Release Notes
## What's Changed * feature: add support to skip jobs that don't have modified files in CI by @ndizazzo in https://github.com/michaelneale/mesh-llm/pull/116 * launch: asymmetric KV cache — Q8_0 K + Q4_0 V for 5-50GB models by @michaelneale in https://github.com/michaelneale/mesh-llm/pull/104 * Fix compilation errors from missing `served_model_descriptors` in test struct initializers by @Copilot in https://github.com/michaelneale/mesh-llm/pull/120 * Fix compile errors in model descriptor test helpers by @Copilot in https://github.com/michaelneale/mesh-llm/pull/121 * Fix UI build to fail on TypeScript compile errors by @i386 in https://github.com/michaelneale/mesh-llm/pull/124 * Add runtime model load/unload with CLI and REST by @i386 in https://github.com/michaelneale/mesh-llm/pull/56 * URGENT: Fix client --auto join hang (connect_to_peer regression) by @michaelneale in https://github.com/michaelneale/mesh-llm/pull/126 * ci: add client-auto boot test by @michaelneale in https://github.com/michaelneale/mesh-llm/pull/125 * Add canonical served model descriptors by @i386 in https://github.com/michaelneale/mesh-llm/pull/118 * Decomposing the codebase by @i386 in https://github.com/michaelneale/mesh-llm/pull/127 * Remove Windows builds from CI by @i386 in https://github.com/michaelneale/mesh-llm/pull/140 * router: context-aware host selection and retry on context overflow by @michaelneale in https://github.com/michaelneale/mesh-llm/pull/141 **Full Changelog**: https://github.com/michaelneale/mesh-llm/compare/v0.54.0...v0.55.0