Ollama v0.32.15 was recently released, delivering significant performance enhancements and bug fixes for local large language model inference. A primary improvement involves caching resolved model metadata between requests, which the release notes state has "cut time-to-first-token by roughly half." Benchmarks provided indicate that time-to-first-token dropped from approximately 995 milliseconds to 524 milliseconds. The update also resolves a bug where chat and generate functions could become unresponsive following a mid-stream parser error. System messages for Qwen 3.8 models are now normalized to ensure consistent handling of non-leading system messages. The release incorporates dependency updates for MLX and llama.cpp, alongside a new desktop onboarding flow for first-time users.