Views
No views yet
google/gemma-4-12B-it using oQ mixed-precision quantization.processor_config.json file was corrected so current MLX-VLM/oMLX Gemma 4 unified loading can use the local Gemma4UnifiedProcessor instead of falling back to the Transformers processor path. The previous metadata passed unsupported video_processor keys (do_convert_rgb, do_sample_frames, resample, return_metadata) into Gemma4UnifiedVideoProcessor. In oMLX this could make VLM loading fail, after which oMLX could silently fall back to a text-only LLM engine. In that fallback state image, audio, and video inputs do not reach the model even though text generation still works.vlm.input_audio with WAV base64 was verified locally after this processor metadata fix. Keep audio clips within the upstream Gemma 4 12B limit of 30 seconds.video / input_video content parts before they reach the VLM engine. For oMLX today, use extracted frames/contact sheets as images, or use a runtime/API path that explicitly supports video preprocessing./v1/models/status shows this model as engine_type: "batched" or model_type: "llm", multimodal inputs are not being used. Restart/reload oMLX after updating this file and confirm the model is routed as vlm.curl http://127.0.0.1:8000/v1/models/status1{
2 "id": "gemma-4-12B-it-oQ4",
3 "engine_type": "vlm",
4 "model_type": "vlm",
5 "config_model_type": "gemma4_unified"
6}