Google Launches Gemma 4 While US Government Explores Strategic Stake in OpenAI
Large Language Models
Google Releases Gemma 4 with Quantization-Aware Training: Google has launched Gemma 4 featuring quantization-aware training (QAT), which allows lower precision models to maintain high performance levels. This advancement enables Q4 quantized models to perform similarly to Q8 models, significantly increasing efficiency for deployment on resource-constrained edge devices.
DeepSeek V4 Flash Integrated into llama.cpp: The new DeepSeek V4 Flash model is being integrated into the llama.cpp framework, utilizing a hybrid FP4-FP8 precision architecture. Early testing indicates the model provides superior intelligence and efficiency compared to other models in its size class.
AI Agents & Frameworks
OpenLumara Framework Launched for Local AI Agents: OpenLumara is a modular, token-efficient AI agent framework built from scratch specifically for local model execution. It features a minimal system prompt and enhanced security measures while supporting complex tasks like web browsing and coding.
AI Services & Products
Perplexity AI Adds NVIDIA Nemotron 3 Ultra: Perplexity has integrated NVIDIA’s Nemotron 3 Ultra model into its platform for Pro and Max subscribers. This model is specifically designed to power long-running AI agents and enhances the platform's advanced search and computational capabilities.
Industry & Policy
US Government Discussing Stake in OpenAI: The Trump administration is reportedly in discussions with OpenAI regarding a potential government stake in the startup. This move suggests a major shift toward direct government investment and involvement in the strategic development of national AI technologies.