“Brad Gerstner has heard that Meta's Llama 4 will be a 400 billion parameter Mixture-of-Experts model with a 10 million token context window and is expected to launch this month.”