llama.cpp release b9833 adds native MiniCPM5 tool call parser
New b9833 enables tool calling for MiniCPM5 models on local hardware.
The b9833 release of llama.cpp, the popular local LLM inference engine, introduces comprehensive support for MiniCPM5 models, including a new tool call parser. This update implements a dedicated MiniCPM5 PEG (Parsing Expression Grammar) parser, refactors the shared mapper to remove streaming tool-arg placeholder workarounds, and adds a jinja2-style template file (openbmb-MiniCPM5-1B.jinja). Key fixes address grammar triggers for preserved tokens and improve streaming tool-argument handling. The autoparser now supports XML tool calls specifically for MiniCPM5, while generic parsing remains strictly JSON-based.
Builds are distributed across all major platforms: macOS (Apple Silicon and Intel), Linux (x64/arm64 CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (x64/arm64 CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Android (arm64), and openEuler. Some configurations are disabled by default (macOS Intel KleidiAI, openEuler). This release makes MiniCPM5 with tool calling accessible to developers running LLMs locally, expanding the ecosystem beyond cloud-dependent solutions.
- Implements a dedicated MiniCPM5 PEG parser for tool call parsing and grammar handling
- Refactors streaming tool-arg placeholders and restores strict JSON parsing for generic mapper
- Builds available for macOS, Linux, Windows, and Android with diverse GPU backends
Why It Matters
Brings advanced MiniCPM5 tool calling to local LLM inference, enabling offline AI agent workflows on consumer hardware.