“Brad Gerstner has heard that Meta's upcoming Llama 4 model will be a 400-billion parameter Mixture-of-Experts model with 50-70 experts, a 10 million token context window, and is expected to launch this month.”