The open-source Ollama project released version 0.33.3 of its local AI runtime on 2 September 2026. The headline change brings multimodal Gemma 4 to Apple’s MLX engine: on Apple Silicon the model now handles images and audio in addition to text, directly on the device.
Multimodal on the Mac
The release notes put it tersely: “gemma4 now supports images and audio on MLX engine.” That means Google’s open Gemma 4 model, including image and audio input, now runs on Apple’s MLX runtime – no detour through CUDA GPUs required. For developers on MacBooks or Mac Studio machines, multimodal prompts run locally, without the cloud and without a dedicated Nvidia card.
More Honest Token Accounting
The project lists two further concrete changes: Ollama now reports cached prompt tokens separately and honors the default parameters a model defines inside its GGUF file. Both make runs easier to reproduce and give a more honest picture of cost and context-length usage.
- Gemma 4 with images and audio on the MLX engine
- Reporting of cached prompt tokens
- Respect for GGUF-defined default parameters
- Updated MLX, MLX-C and llama.cpp engines
The 0.33 series had already reworked the internal engine bindings. Version 0.33.3 refreshes the embedded MLX, MLX-C and llama.cpp components, continuing the roughly weekly release cadence the team has kept up since the summer.
Sources: GitHub – Ollama v0.33.3 · Ollama Releases
