llama.cpp v1.0.0, First Major Release in Three and a Half Years

• The official v1.0.0 release was announced approximately three and a half years after the first commit in March 2023
• A semantic versioning policy was introduced that increments the major version if public C API compatibility is broken
• 35 functions in llama.h were removed or renamed, and the llama-server security policy was strengthened
• Support for FP8 KV cache and new model architectures such as Kimi-K3.1-Pro and Ripple-65B-Preview was added

Compatibility and API Changes

• Existing function name change: llama_load_model_from_file was changed to llama_model_load_from_file
• 18 functions related to NextN and speculative decoding from llama-ext.h were moved to llama.h
• The ggml RPC protocol was upgraded to v9, and connections are rejected upon version mismatch

Features and Performance Improvements

• KV cache storage in OCP FP8 format is now possible with the -ctk f8_e4m3 option
• Parameters related to MoE expert caching were added, allowing the CPU processing ratio to be specified
• Quantized matrix multiplication and MoE fusion performance were improved in major backends such as CUDA, Metal, and Vulkan

Distribution and Build

• Version 1.0.0 is available on Homebrew and MacPorts, while GitHub and Docker are based on nightly b12359
• COMPATIBILITY SameMajorVersion is applied in the CMake configuration, allowing only 1.x versions
• The default for development builds is 1.0.0-dev, and specifying an option is required to use the official version string