b11115 v0.5.0
ggml-org/llama.cppSep 22, 2026 → Sep 23, 20261 day between releases
What's New
Introduces CUDA/Metal optimizations, new models, and multi-address server binding.
Breaking Changes
- The `--host` flag now accepts comma-separated TCP addresses and UNIX sockets, which may break existing single-address configurations.
New Features
- Accelerate CUDA `conv2d` with implicit GEMM.
- Metal MoE and SSM_CONV fusion optimizations.
- Multi-address binding support via `--host`.
- New model support for HRM-Text, MiMo-V2.6, and HunyuanOCR.
- Environment variables for temperature, top-p, min-p, and penalties.
Improvements
- Updated ggml to v0.25.0 with expanded hyper-connection and flash-attention support.
- Accelerate CUDA `conv2d` with implicit GEMM.
- Metal MoE and SSM_CONV fusion optimizations.
- Server robustness with multi-address binding.
- HunyuanOCR conversion support.
Bug Fixes
- Several chat parser and UI fixes implemented.
Upgrade Notes
Update your `--host` configuration to use comma-separated addresses if binding to multiple interfaces. New environment variables for chat parameters are now available. Be aware that `llama_model_load_from_file_ptr()` now requires aligned mmap.
The Two Releases
b11115
Sep 22, 2026· 29 days ago
Adds a new OpenCL binary kernel for A8 Q4_K non-MoE operations and refactors helper functions to improve kernel selection logic.
Full release notesv0.5.0
Sep 23, 2026· 28 days ago
This release prioritizes backend performance and correctness, introducing support for new models like HRM-Text and MiMo-V2.6, and significantly improving server robustness with multi-address binding capabilities.
Full release notes