Ollama v0.40.12 Released: Adds Liquid AI LFM3 and d1 Models

Ollama v0.40.12 Released: Adds Liquid AI LFM3 and d1 Models

At a Glance

Item Value
Repository ollama/ollama
Version v0.40.12
Published 2026-10-27
License MIT
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Ollama has released its latest version, v0.40.12. Ollama is a tool designed to easily run various open-weight models, such as Gemma, DeepSeek, and Qwen, in local environments. This update adds support for Liquid AI’s new model family, “LFM3" and “d1 decision models," and brings faster prompt processing for specific models like Gemma 4, particularly in Apple Silicon environments.

Breaking Changes and Deprecations

None.

Key Changes

Support for the Liquid AI LFM3 Family

Support has been added for the “LFM3" model family developed by Liquid AI, designed for local agents, tool calling, and structured outputs. This family includes lfm3, a 1.7B text model with a 256K token context window; lfm3-vl, which adds image understanding; lfm3-audio, capable of speech recognition, translation, and text-to-speech across 18 languages; and lfm3-omni, which handles text, images, and audio in a single model.

Note that while this release supports running lfm3 and d1 2b on Apple hardware, support for Linux and Windows is scheduled to arrive soon.

Addition of d1 Decision Models (2B)

The 2B size of Liquid AI’s d1 model, built on the LFM3 architecture, has been added. d1 makes decisions regarding text and images, while d1-omni also supports audio. For example, it can be run using a command like the following:

ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"

Faster Prompt Processing on Apple Silicon

For users running Apple Silicon, prompt processing speeds have been improved for the Gemma 4, Flint, and Laguna models. Users running these models on Apple chips such as Macs can expect smoother operation following this update.

Inference Improvements and Bug Fixes

Improvements have been made to inference performance and behavior in specific environments. DFlash2 draft models are now supported for speculative decoding in MLX environments. Additionally, Gemma 4’s ability to retain detailed information contained in small images has been improved. Furthermore, ollama launch now works with OpenCode v2, and fixes related to tool-only assistant turns in Kimi CLI are also included. These changes accompany internal updates to MLX and llama.cpp.

Supported Models and Hardware

With this update, the following model architectures and models are now available:

Liquid AI LFM3 Series

The LFM3 family, based on a new “tri-hybrid architecture" developed by Liquid AI, is supported. These models are designed for local agents, tool calling, and structured outputs.

  • lfm3: A text model with 1.7B parameters. It supports a 256K token context window.
  • lfm3-vl: A model adding image understanding capabilities to lfm3.
  • lfm3-audio: A model supporting speech recognition, translation, and text-to-speech in 18 languages.
  • lfm3-omni: A multimodal model capable of handling text, images, and audio in an integrated manner within a single model.

d1 Decision Models

The 2B size of the d1 model, specialized for decision making and built on the LFM3 architecture, has been added.

  • d1: A model that makes decisions regarding text and images.
  • d1-omni: A model adding audio capabilities to d1.

Supported Hardware and Optimizations

Regarding hardware and runtime environments, the following points are reported:

  • Apple Hardware: Running lfm3 and d1 2b is supported on Apple hardware in this release. In addition, prompt processing for Gemma 4, Flint, and Laguna models has been accelerated in Apple Silicon environments. Furthermore, Gemma 4 has improved capability in retaining detailed information within small images.
  • MLX Environments: DFlash2 draft models are newly supported in speculative decoding.
  • Other Platforms: Support for lfm3 and d1 2b on Linux and Windows environments is reported to be coming soon.

How to Get It

To update Ollama to the latest version, obtain and apply the latest installer from the official releases page.

When using new models, use the standard ollama run command. For example, to run lfm3:

ollama run lfm3

Additionally, to run instructions combining text and questions using the d1:2b model, you can execute it in the following format:

ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"

Releases Since Our Last Article

Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.

Version Released Release notes
v0.40.11 2026-10-24 GitHub
v0.40.10 2026-10-24 GitHub

Related Articles

What to Read Next

Sources