Ollama v0.32.1 Fixes MLX Memory Leak, Improves Gemma 4 Tool Calling

Ollama released version 0.32.1 on July 16, 2024, addressing a recurrent memory leak in MLX model caching and improving Gemma 4's tool calling capabilities.

Key Fixes

The release resolves a critical MLX model cache leak that caused memory usage to increase across requests. According to the changelog, the fix also improves cache snapshot performance for Apple Silicon users running models through the MLX framework.

MLX text model loading now respects the OLLAMA_LOAD_TIMEOUT environment variable, giving users more control over model initialization timeouts.

Gemma 4 Improvements

Ollama claims improved Gemma 4 tool calling and multi-turn reasoning, specifically citing "more reliable tool-response continuations." The changelog does not specify benchmark improvements or quantify the reliability gains.

Agent and Authentication Updates

The agent web search and fetch features now prompt users to run ollama signin when authentication is required, rather than failing silently. The interactive agent receives the current working directory for improved project context awareness.

Developer Experience

The update fixes ollama launch behavior when a deprecated model is passed with the --model flag—the model picker now opens correctly instead of erroring. Documentation for the official Ollama VS Code extension setup has been updated.

What This Means

The MLX memory leak fix is significant for Mac users running local models, as uncontrolled memory growth could cause system instability during extended inference sessions. The Gemma 4 tool calling improvements suggest Ollama is optimizing for agentic workflows, though without specific benchmarks, the actual reliability gains remain unclear. This patch release indicates active maintenance of platform-specific issues, particularly for Apple Silicon optimization.