Keep pulling the thread on Omar.
The Jemma model family achieved over 40 million downloads within the first three weeks of its launch.
Google changed the license for the Jemma 4 models to Apache 2.0 in response to community feedback that the previous custom license was a friction point for adoption.
The Jemma 4 launch was Google DeepMind's largest open model release to date.
The Jemma family of models ranges in size from 2 billion to 31 billion parameters.
The smallest Jemma models are capable of running on a phone.
The largest Jemma model, at 31 billion parameters, is small enough to run on a consumer-grade GPU.
The smallest Jemma models are multimodal and can understand audio, video, and image inputs.
The Jemma models were trained on data from over 140 languages.
The open-source community is fine-tuning Jemma models for specific language tasks, such as improving Quechua to Spanish translation.
Jemma models have agentic capabilities, including function calling and the ability to use different APIs to solve tasks.
The open-source community is successfully deploying Jemma models on hardware-constrained devices such as Raspberry Pis and Jetson Nanos.
Omar from Google DeepMind estimates that 70-80% of daily user tasks could potentially be fulfilled by local AI models.