Llama.cpp: now with automatic parser generator

The open-source llama.cpp project has successfully merged a groundbreaking 'autoparser' system into its mainline code after months of testing and refinement. This novel architecture, built on top of recent foundational changes like ngxson's native Jinja templating system and aldehir's PEG (Parsing Expression Grammar) parser, aims to automatically generate the correct parser for most AI model templates. By analyzing the common patterns models use to define reasoning, tool calls, and content parsing, the autoparser can infer the necessary logic directly from a model's template file. This eliminates the need for developers to manually write and maintain custom parsers for each new model that follows standard conventions, a significant quality-of-life improvement for the ecosystem.

The autoparser handles the majority of cases, but for models with unique or overly complex formats—such as GPT OSS's Harmony or Kimi 2.5's specific function-calling syntax—developers can still write custom parsers using the unified PEG framework. A workaround system also exists for legacy models. This centralized approach replaces a patchwork of individual parsers with a systematic solution, making llama.cpp more reliable for 'agentic work' where AI agents take actions. An imminent related update will fix a nagging issue for Qwen 3.5 models getting stuck in `read_file` loops, further cementing llama.cpp's role as a stable backbone for running and experimenting with various open-source and proprietary AI models locally.