M-Courtyard v0.5.3: Better Gemma 4 Local Training, Improved Ollama Export, and Stability Fixes

We’re happy to share M-Courtyard v0.5.3.

This release is focused on improving the real-world local fine-tuning experience on Apple Silicon, with special attention to Gemma 4 support , training reliability , and export stability.

What’s improved in v0.5.3

Better Gemma 4 local training support

M-Courtyard now handles Gemma 4 local model layouts more reliably, improving local model detection and making the training workflow smoother for Gemma 4 users.

Improved Ollama export compatibility

We added better Gemma 4 compatibility in the export pipeline, making it easier to turn trained models into Ollama-ready packages.

macOS environment setup fixes

This release fixes a regression in the macOS environment setup flow, improving reliability when preparing the local training environment.

General bug fixes and stability improvements

v0.5.3 also includes multiple fixes across the local fine-tuning workflow to make training and export more dependable in daily use.

About M-Courtyard

M-Courtyard is a privacy-first desktop app for Apple Silicon that helps users go from raw text to fine-tuned local models with a guided local workflow.

Release

GitHub release:

GitHub

Release M-Courtyard v0.5.3 · Mcourtyard/m-courtyard

v0.5.3 Fixed macOS uv --system-certs regression Gemma 4 local model detection and training Gemma 4 Ollama export compatibility Changed Online model lists: Qwen 3.5, Gemma 4, removed Phi mlx-lm[t...