ollama

Key Features

• Liquid AI LFM3 model family added: New models based on a tri-hybrid architecture for local agents, tool calling, and structured output.
  • lfm3: 1.7B text model with 256K token context window support.
  • lfm3-vl: Added image understanding capabilities.
  • lfm3-audio: Supports speech recognition, translation, and text-to-speech conversion in 18 languages.
  • lfm3-omni: Processes text, images, and audio in a single model.
• d1 decision models 2B version added: Support for d1, a text and image decision-making model based on LFM3, and d1-omni, which adds audio support.
• Speculative decoding support for DFlash2 draft models in MLX environments.
• ollama launch command updated for compatibility with OpenCode v2.

Performance Improvements

• Improved prompt processing speed for Gemma 4, Flint, and Laguna models in Apple Silicon environments.
• Improved Gemma 4 to retain more detail in small images.

Bug Fixes

• Fixed tool-only assistant turn handling error in Kimi CLI.

Other

• Updated MLX and llama.cpp libraries.
• Note: lfm3 and d1 2b models support Apple hardware only in this release; Linux/Windows support is planned.