Llama 4 is lost

Llama 4 has been met with a mixed reception following its early release by Meta AI. Two versions—Scout and Maverick—were launched ahead of schedule, briefly gaining attention for their performance on the LMArena leaderboard. However, closer inspection revealed that the benchmarked version of Maverick was specifically optimized for that test and visual diagrams published by Meta AI excluded competing models like Gemini 2.5 Pro, which outperforms it.

The launch comes a year after Llama 3 was widely seen as the open standard in AI, but since then Meta AI has faced internal turbulence, including the departure of key figures like Joelle Pineau, their Head of AI Research.

The broader open-source AI landscape has shifted significantly in the meantime. Models such as DeepSeek V3 and R1, along with Alibaba’s Qwen-2.5-Max and Qwen-QwQ, now outperform Llama 4 in areas like coding, and are released under far more permissive licenses (MIT and Apache 2.0). In contrast, Llama 4’s license includes branding requirements and usage restrictions, including a clause barring companies with over 700 million monthly active users.

These limitations have drawn criticism, including from OpenAI CEO Sam Altman, who noted that their upcoming open model will not include similar restrictions.