September 2025 Summaries

Ollama has launched a new web search API that enhances its models, such as OpenAI's gpt-oss, by incorporating the latest web information to improve accuracy and reduce hallucinations. Available as a REST API, it offers a free tier for individuals and higher rate limits through its cloud service. The API integrates with Ollama's Python and JavaScript libraries, enabling long-term research tasks and the creation of mini search agents. Users can initiate the API by generating an API key and making requests via cURL, Python, or JavaScript. Additionally, the platform supports fetching individual pages through a separate web fetch API, and it integrates with various systems like MCP Server, Cline, Codex, and Goose, extending its usability across different applications.

Ollama has introduced an enhanced model scheduling system that precisely measures memory requirements before executing a model, improving upon the previous estimation-based approach. This advancement reduces out-of-memory crashes by preventing over-allocations and maximizes GPU utilization by allocating more memory to the GPU, thereby increasing token generation and processing speeds. Additionally, the new system optimizes performance across multiple GPUs, enhancing multi-GPU and mismatched GPU operations, and ensures accurate memory utilization reporting, aligning with tools like nvidia-smi. All models on Ollama's new engine now feature this memory management improvement by default, with more models transitioning soon, resulting in significant performance gains as demonstrated with NVIDIA GeForce RTX 4090 GPUs, showing increased token generation and prompt evaluation speeds across supported models.

Ollama has introduced cloud models in preview, enabling users to run larger AI models on datacenter-grade hardware without compromising on privacy, as the cloud does not retain user data. These models integrate seamlessly with local tools and the existing Ollama experience, allowing users to execute large-scale models that might not be feasible on personal computers. Users can interact with these cloud models through an OpenAI-compatible API, employing languages like JavaScript, Python, and cURL to perform tasks such as pulling, running, and managing models. The Ollama cloud requires users to sign in for accessing cloud models but also offers an option to sign out if preferred.