Ollama 0.40.12 adds Liquid AI LFM3 and d1 models, Apple Silicon optimizations

Ollama version 0.40.12 introduces support for Liquid AI's LFM3 model family and d1 decision models, expanding the local AI runtime's multimodal capabilities. The LFM3 family includes a 1.7B-parameter text model with a 256K-token context window, plus vision-language, audio, and omni variants that handle text, images, and audio in a single model. The d1 decision models, now available in a 2B size built on LFM3, are designed for making decisions about text and images, with d1-omni adding audio processing.

However, both LFM3 and d1 2B models are restricted to Apple hardware in this release, with Linux and Windows support explicitly marked as coming soon. The update also delivers faster prompt processing on Apple Silicon for Gemma 4, Flint, and Laguna models; adds speculative decoding support for DFlash2 draft models via MLX; improves Gemma 4 image detail retention; fixes compatibility with OpenCode v2 and Kimi CLI tool turns; and updates the underlying MLX and llama.cpp dependencies.

What changes in Ollama 0.40.12

• Liquid AI LFM3 integration: New model family built on a tri-hybrid architecture, offering 1.7B parameters with 256K-token context, plus multimodal variants for vision (lfm3-vl), audio in 18 languages (lfm3-audio), and combined omni processing (lfm3-omni).
• d1 decision models: 2B-parameter models built on LFM3 for decision-making tasks over text and images, with d1-omni extending to audio; invoked via ollama run d1:2b with --ask flag for structured queries.
• Apple Silicon performance: Faster prompt processing for Gemma 4, Flint, and Laguna models; speculative decoding now supports DFlash2 draft models through the MLX backend.
• Gemma 4 image handling: Improved retention of visual detail when processing small images.
• Infrastructure updates: ollama launch now compatible with OpenCode v2; fixed tool-only assistant turns from Kimi CLI; MLX and llama.cpp dependencies refreshed.

Why it matters

The LFM3 integration signals growing commercial interest in Ollama as a distribution channel for proprietary local AI models, but the Apple-only availability creates a platform gap that leaves Linux and Windows developers waiting. For the broader open-source AI tooling landscape, the continued MLX and llama.cpp updates maintain Ollama's position as a consolidation layer for local inference backends, though users should note that Liquid AI's model weights and architecture remain proprietary despite Ollama itself being open source.