Google DeepMind

Google DeepMind is an AI research lab formed from the 2023 merger of Google Brain and DeepMind. It develops large language models (e.g., Gemini family) and multimodal systems. For local AI operators, Google DeepMind's models are relevant when they are released as open-weight (e.g., Gemma 2) or when their architecture influences open-source implementations. The lab's research on efficient training and distillation often trickles down into smaller, runnable models.

Google DeepMind focuses on advancing AI through fundamental research, reinforcement learning, and large-scale neural networks. Its most notable contributions include the Transformer architecture (via Google Brain), AlphaFold for protein folding, and the Gemini series of multimodal models. While many of its flagship models are proprietary, Google DeepMind has released open-weight models like Gemma 2 (2B, 9B, 27B) under permissive licenses. These models can be run locally with tools like llama.cpp or Ollama. The lab's research on mixture-of-experts (MoE) and quantization-aware training directly impacts the efficiency of models that operators can deploy on consumer hardware.

An operator running ollama pull gemma2:9b downloads a 9B-parameter model developed by Google DeepMind. On an RTX 4090 (24 GB VRAM), the Q4 quantized version (~5.5 GB) runs at ~40 tok/s. The model's architecture includes grouped-query attention, which reduces memory bandwidth requirements compared to full multi-head attention, making it more efficient on consumer GPUs.

When using Hugging Face Transformers, an operator might load a Gemma model with AutoModelForCausalLM.from_pretrained('google/gemma-2-9b-it'). The tokenizer and config are specific to Google DeepMind's design. For local inference, llama.cpp supports Gemma models via GGUF format; the operator converts the model using convert_hf_to_gguf.py and runs it with ./main -m gemma-2-9b-Q4_K_M.gguf -p 'Hello'.