Ollama v0.40.12: LFM3 and d1 2B models

Summary

Ollama 0.40.12, released on 2026-10-27, adds Liquid AI's LFM3 family of small models and a 2B size of the d1 decision models built on LFM3. It also adds DFlash2 speculative decoding on MLX, faster prompt processing on Apple silicon, and ollama launch support for OpenCode v2.

Key points

• LFM3 (Liquid AI): lfm3 is a 1.7B text model with a 256K-token context window. Variants: lfm3-vl (images), lfm3-audio (speech recognition, translation, and text-to-speech in 18 languages), and lfm3-omni (text, images, and audio).
• d1 2B: a new size of the d1 decision models, built on LFM3; d1 handles text and images, d1-omni adds audio.
• Apple only for now: LFM3 and d1 2B run only on Apple hardware in this release; Linux and Windows support is planned.
• Speed: DFlash2 draft models for speculative decoding on MLX; faster prompt processing for Gemma 4, Flint, and Laguna on Apple silicon; better small-image detail in Gemma 4.
• Try it: ollama run lfm3 (or lfm3-omni). ollama launch now works with OpenCode v2.
• The release notes do not state LFM3's licence (checked 2026-10-28).

Why it matters

A 1.7B model with 256K context and an all-in-one audio and vision variant is small enough for laptops and edge devices.