llama.cpp adds reasoning_effort for OpenAI compatibility
llama.cpp v1.3.0 now supports OpenAI's reasoning_effort parameter for structured outputs
Deep Dive
The llama.cpp repository released build b10434, adding support for OpenAI Chat Completions reasoning_effort and exposing it through chat templates and generation parameters. The update includes related fixes and is available for download across multiple platforms.
Key Points
- llama.cpp v1.3.0 adds OpenAI reasoning_effort parameter for controlling model reasoning depth
- Pre-built binaries available for 18+ platforms including Windows, Linux, macOS, Android, and openEuler
- Supports CUDA 12/13, Vulkan, ROCm, OpenVINO with GPU acceleration across multiple architectures
Why It Matters
Enables precise control over LLM reasoning for local deployments while maintaining OpenAI API compatibility