Google DeepMind Releases Gemma 4: Open-Weight Multimodal Models Running on a Single H100 GPU with 140+ Language Support

Google DeepMind launched the Gemma 4 family, four open-weight multimodal models (Apache 2.0 licensed) engineered to run entirely on a single 80GB Nvidia H100 GPU while delivering benchmark-competitive performance across image, text, and audio inputs. The family includes Gemma's first Mixture-of-Experts model, supports over 140 languages, and is optimized for reasoning, coding, agentic workflows, and on-device deployment from smartphones to IoT systems — reducing latency and cloud dependency. Released simultaneously on Hugging Face with NVIDIA-optimized variants, Gemma 4 enters a crowded open-model field alongside Meta's Llama 4 Scout (10M token context), Alibaba's Qwen 3.6-Plus (1M token context), DeepSeek, and others. Developers can also run Gemma 4 locally using LM Studio's headless CLI interface, a workflow that earned 310 points on Hacker News.