Meta has launched its Llama-4 AI models, which promise to compete with top closed-source models without fine-tuning. The flagship model, Llama 4 Behemoth, has 288 billion parameters and purportedly outperforms existing models like GPT-4.5 and Gemini Pro on STEM benchmarks. The new models, Llama 4 Scout and Maverick, each utilize 17 billion active parameters but differ in their expert configurations, enabling efficient processing on less powerful hardware. Despite bold claims about capabilities, some independent tests revealed inconsistencies in performance, particularly with longer context prompts and logical reasoning tasks. For instance, when tested with complex inquiries, Llama 4 struggled to retrieve exact information and produced subpar results in logical puzzles. Users also noted that Llama 4 exhibited a verbosity that could hinder casual conversation. In terms of creative writing, it showed promise, generating immersive narratives but lagging behind competitors in emotional depth. While the models hold potential as open-source bases for further tuning, they require substantial improvements to meet Meta's lofty claims. Overall, Llama 4 marks a significant step in the AI landscape, though it has a way to go before achieving its marketed status.

Source 🔗