Ollama v0.34.4 was released today, bringing several enhancements for local large language model inference. The update improves structured outputs on thinking models, making them faster and more reliable through a single-pass application. It also addresses intermittent "model not found" errors that occurred with large local model libraries and fixes an issue where the macOS application became unresponsive. For Apple Silicon users, Qwen 3.8 prompt processing is now faster, and Gemma 4 models gained improved image resolution selection, ensuring "more detail in high-resolution images." The release further includes updates to underlying components such as llama.cpp, MLX, and XGrammar, continuing the evolution of Ollama's runtime capabilities.