The latest release brings a significant boost to Ollama’s experimental agent features. A new skills system allows agents to leverage predefined capabilities, and cloud models now enjoy unlimited tool rounds by default. Command-line improvements include proper slash command completion before submission, removal of redundant code, and cleaner server start-up routing. The terminal user interface also gets a polish with consistent bold markdown rendering across chat outputs.
Under the hood, the llama.cpp library has been updated to incorporate upstream improvements, and the MLX backend for Apple Silicon sees an update. The Anthropic provider integration now correctly closes text blocks before opening thinking blocks, ensuring smoother AI interactions. The Claude Code channels remain available, and the Hermes integration has been refreshed.
For self-hosted users on Linux, the build toolchain has been bumped to GCC 13, and CUDA v12 now supports compute capability 10.0, extending compatibility with newer NVIDIA GPUs.
