Ollama Adds Audio Generation and New Decision Models

What happened

Ollama has added speech generation capabilities using Phonon on Apple Silicon devices, allowing users to generate and play audio in real-time with the 'ollama run phonon' command. The release also includes an OpenAI-compatible endpoint for streaming audio through the '/v1/audio/speech' route.

The update introduces d3 decision models, which are derived from vLLM Semantic Router's decision models, and are available in sizes of 2B, 4B, 9B, and 27B parameters. Flint, a tool for interacting with models, now supports Linux and Windows in addition to macOS.

Why it matters

These updates expand Ollama's capabilities in audio generation and decision-making, supporting more use cases for developers and users. The improved compatibility with coding tools and web search features enhances the reliability of model interactions.

Key details

The d3 models can classify, answer yes/no questions, or score options with probabilities based on input text, structured state, or images. The 'ollama launch' command now correctly transfers cloud model context lengths and capabilities to coding tools, resolving previous issues.

The release includes a full changelog at the specified GitHub link for version v0.40.9. A sample command is provided: 'ollama run phonon "Hello from Ollama."'.