v0.18.2-rc0

Ollama, the open-source platform for running large language models locally, has pushed a new release candidate: version 0.18.2-rc0. This pre-release focuses on performance enhancements and memory management, particularly for users on Apple Silicon hardware. The most significant change is the introduction of a model eviction scheduler for the MLX framework, Apple's machine learning library for its custom silicon. This allows Ollama to better manage system memory by unloading inactive models, a crucial feature for running multiple or larger models on devices with limited RAM.

Other technical improvements include adding prequantized tensor packing specifically for the Qwen 3.5 model family, which should speed up loading and inference times. The update also includes fixes for quantized embeddings and the SwiGLU activation function within the MLX backend. For users of the web search feature, the release adds a fix to flush output on newlines for the legacy path and registers the feature for the 'openclaw' model. These are incremental but important optimizations that refine the experience of running state-of-the-art open-source models like Qwen 3.5 on personal computers.