Ollama v0.33.3-rc2 enables image and audio processing for Gemma 4 Safetensors imports running on the MLX engine1. Images work across both vision architectures: the transformer tower (26B, 31B, e-series) and the 12B's encoder-free unified embedder. Audio input uses the same API paths already available for Gemma 4 GGUFs, including WAV bytes, OpenAI-compatible input_audio parts, and transcription uploads. The e2b/e4b checkpoints process audio through their conformer encoder, while the 12B unified checkpoint embeds raw waveforms directly. Clips exceeding 30 seconds are split at pauses into chunks of at most 30 seconds.
Ollama Adds Image and Audio Input for Gemma 4 Safetensors via MLX Engine
Ollama v0.33.3-rc2 enables image and audio input for Gemma 4 Safetensors imports served by the MLX engine, supporting multiple vision and audio…
