# Ollama 0.40.12-rc1 adds LFM3 and d1 model support via Apple MLX backend

> **Key takeaway:** Ollama's release candidate expands Apple Silicon local AI inference by adding LFM3 and d1 model architectures to its MLX backend.

**Published:** 2026-10-27T15:47:18+00:00 
**Source:** Ollama Local AI Runtimes 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/ollama-04012-rc1-adds-lfm3-and-d1-model-support-via-apple-ml.html 

## Summary
Ollama has published v0.40.12-rc1, a release candidate that extends its Apple MLX (Machine Learning eXtensions) backend with support for two additional model families: LFM3 and d1. These additions build on Ollama's existing ability to run large language models locally on Apple Silicon hardware through Apple's native ML framework. The release continues Ollama's pattern of rapid iteration on its multi-backend runtime, which supports CUDA, ROCm, Vulkan, and CPU inference in addition to MLX.

Ollama is distributed under the MIT license and provides a Go-based inference server with pre-built binaries for macOS, Linux, and Windows. The project enables developers to pull, run, and customize quantized models without cloud dependencies. This release candidate, tagged by maintainer Daniel Hiltgen, is available through GitHub releases with signed commits.

## MLX backend and model support changes
- **MLX backend expansion:** The release adds LFM3 and d1 model architectures to Ollama's Apple MLX backend, which leverages Apple's unified memory architecture on M-series chips for efficient local inference.
- **Model family additions:** LFM3 refers to Liquid Foundation Models version 3 from Liquid AI, which use state-space model architectures rather than traditional transformers; d1 likely refers to a distilled or specialized model variant in the same ecosystem.
- **Release packaging:** The rc1 designation indicates this is a pre-release build intended for testing before the stable v0.40.12; assets are distributed through GitHub's release infrastructure with GPG-verified tags.
- **Integration pattern:** The change follows Ollama's established model of adding backend-specific model support through its modular runner architecture, with MLX models typically benefiting from Metal Performance Shaders and unified memory access on macOS.

## Why it matters
For developers and researchers working on Apple Silicon, this release candidate broadens the range of architectures available for fully local, offline inference without cloud API dependencies. The addition of non-transformer architectures like LFM3 through MLX signals growing diversification in deployable model types beyond standard Llama-derived weights, which matters for edge AI experimentation and privacy-preserving applications on macOS and emerging ARM-based Linux systems.