<p>On April 8th, Jinshi Data reported that at 1:30 AM today, Ahmad Al-Dahle, the generative AI leader of Meta Platforms (META.O), published a lengthy article on social media addressing the doubts raised about Llama 4, which was just open-sourced the day before. Ahmad stated that Llama 4 was released as soon as it was developed, so there will inevitably be some differences in model quality across different services. Meta will soon fix these issues to enhance performance. He also denied that pre-training was conducted on the test set. When Meta announced it on their official website, they specifically mentioned DeepSeek, stating that their newly open-sourced Llama 4 Maverick can match the coding capabilities of DeepSeek's newly open-sourced V3 model</p>

META

<p>At 1:30 AM today, Ahmad Al-Dahle, the leader of generative AI at Meta Platforms, published a lengthy article on social media in response to the doubts raised about Llama 4, which was just open-sourced the day before yesterday. Ahmad stated that Llama 4 was released as soon as it was developed, so there will inevitably be some differences in model quality across different services. Meta will soon fix these issues to improve performance. He also denied that pre-training was conducted on the test set. When Meta announced it on their official website, they specifically mentioned DeepSeek, stating that their newly open-sourced Llama 4 Maverick can match the coding capabilities of its newly open-sourced V3 model</p>

Llama 4 效果不好，Meta 承認有問題