May 2025 Summaries

Ollama introduces a new feature that allows users to enable or disable the "thinking" capability of its models, offering flexibility in how outputs are generated for different applications. When thinking is enabled, the model separates its thought process from the final output, which can be useful for applications requiring detailed reasoning or interactive experiences. Conversely, disabling thinking results in direct and faster outputs. The feature is supported across different interfaces, including the command-line interface (CLI), interactive sessions, and APIs, as well as in Python and JavaScript libraries. This functionality is demonstrated using the DeepSeek R1 model, and the integration allows for versatile application possibilities, such as animating the thinking process in games or other applications.

Ollama has introduced streaming responses with tool calling, allowing chat applications to stream content and call tools in real-time, utilizing models like Qwen 3, Llama 3.1, and Llama 4. This functionality enables users to perform tasks such as fetching weather data or adding numbers through tool calls integrated into the chat interface. The system employs a new parser that focuses on understanding the structure of tool calls rather than relying solely on JSON parsing, which facilitates incremental parsing and improves tool call reliability. Ollama's Model Context Protocol (MCP) further enhances tool calling by allowing expanded context windows, thereby boosting accuracy and performance. The update also includes support for Python and JavaScript, enabling developers to integrate and utilize these tool calls in their applications efficiently.

Ollama has introduced a new engine to support multimodal models, beginning with vision models like Meta Llama 4 and Google Gemma 3, which enhance general multimodal understanding and reasoning. This development allows for sophisticated interactions with visual data, such as interpreting images and answering queries about them, and leverages models like Llama 4 Scout, a mixture-of-experts model with 109 billion parameters. Ollama emphasizes model modularity, ensuring each model can operate independently to improve reliability and ease integration for developers, while also focusing on accuracy and memory management through methods like image caching and optimized KV cache usage. Collaborations with hardware manufacturers and software partners aim to enhance memory efficiency and support longer context sizes, enabling more robust use of multimodal capabilities. This initiative sets the foundation for future expansions in modalities like speech and video generation, with contributions from partners like GGML and major tech companies such as NVIDIA and Microsoft.