Google's Gemma 4 12B model faces criticism for slow local performance and poor multimodal generation on Apple Silicon.

Google's new encoder-free multimodal architecture may face performance penalties when running locally on consumer hardware.

* A speed of 44 tokens/second on an M5 Max with 128GB unified memory is slow for a 12B parameter model, failing to meet developer expectations for local execution.

* The poor quality of the generated lava lamp suggests that the model's multimodal capabilities are not yet competitive with other options.

* The recommendation of Qwen 3.6 27B or 35B highlights Qwen's strong position in the open-weights space, particularly for local deployment.