llama.cpp b11020 Adds Message Delimiters for DeepSeek V3.2/V4 Chat Parsing
llama.cpp release b11020 updates its chat parser to include message delimiters specifically for DeepSeek V3.2/V4 models. This update was co-authored by Sigbjørn Skjæret from Hugging Face and assisted by Claude. Proper chat template parsing ensures that large language models process multi-turn conversations accurately without misinterpreting role boundaries or special tokens. This patch improves compatibility and formatting correctness when running modern DeepSeek model variants locally using llama.cpp. The patch is integrated via pull request #29008 and targets the internal chat parser logic. Pre-compiled binaries for release b11020 have been released across Windows, Linux, macOS, and Android with support for CPU, CUDA, Vulkan, ROCm, and SYCL backends.
## BACKGROUND
llama.cpp is a widely used open-source LLM inference engine optimized for running models locally on consumer hardware. Chat models rely on specific chat templates and special token delimiters to separate system, user, and assistant turns in a prompt. Without correct delimiters, an inference engine might format inputs incorrectly, leading to degraded model outputs or failure to recognize turn endings.