Llama 4 Maverick 400B: First Real-World Reasoning Test

Watch a 17-minute video that evaluates the reasoning capabilities of Llama 4 Maverick, a 400B parameter model with 128 experts and 17B active parameters. Witness a standard logic test conducted on Llama 4 Maverick alongside Claude 3.7 Sonnet Thinking 32K for performance comparison. Discover whether Meta's claim that the model achieves "comparable results to the new DeepSeek v3 on reasoning and coding" holds up under scrutiny. Learn about the testing methodology, which builds upon previously developed causal reasoning tests used for other AI models. The video also mentions that Llama 4 Behemoth (2T parameter) is still in training, while Llama 4 Scout (17B active parameters with 16 experts) can run on a single H100 GPU with Int4 quantization.