v4.5

Oobabooga has launched version 4.5 of its widely-used local AI interface, marking a significant milestone with an official project rename to 'TextGen'. The GitHub repository has moved to github.com/oobabooga/textgen. This release focuses heavily on stability and accessibility, headlined by the introduction of portable, self-contained builds. These standalone packages for Windows, Linux, and macOS allow users to run the software directly after extraction, with no Python or dependency installation required. The builds cater to multiple hardware backends, including NVIDIA CUDA (12.4 and 13.1), AMD ROCm 7.2, Vulkan for AMD/Intel GPUs, and CPU-only operation, significantly lowering the barrier to entry for running local models like Llama 3 or Mistral.

Under the hood, v4.5 delivers crucial bug fixes that improve the experience with newer model architectures. It resolves Gemma-4's tool-calling feature, which was failing to handle double quotes and newline characters in arguments. A memory optimization reduces VRAM peak usage during the prompt logprobs forward pass, which can prevent out-of-memory errors on systems with limited GPU RAM. The update also patches several UI annoyances, such as the chat scroll getting stuck on 'thinking' blocks and incorrect chat deletion confirmations. Dependency updates include pulling in the latest commits from llama.cpp and its fork, ik_llama.cpp, which offers additional quantization types.