What you need to know
- A report suggests Meta’s Llama 4 has encountered a series of internal and external issues, forcing it into multiple delays.
- Sources claim the company experienced increased pressure to do more with its machine learning following DeepSeek’s startling arrival.
- Meta is also looking to improve the AI’s performance, reasoning, and math skills as it chases an alleged launch “later this month.”
- Meta’s CEO Mark Zuckerberg teased Llama 4’s 2025 launch last fall; however, the company also announced LlamaCon for April 29.
A report claims the next iteration of Meta’s LLM is still on the way, though it has run into a few roadblocks.
According to “people familiar with the matter,” The Information (paywall) claims Meta’s Llama 4 model has run into issues that have severely hindered its progress. Allegedly, one of the problems Meta ran into was with its competition, namely the Chinese-based AI model DeepSeek. The publication states these “challenges” concerned the performance of this new competitor and the “additional pressure” to “push the ball forward” with its upcoming AI model.
Sources allege Meta’s Llama 4 model will “borrow” from DeepSeek’s machine-learning approach by deploying a new method called “mixture of experts.”
The post states this new method will split up the AI’s training, focusing on separate parts so each becomes “an expert” in its own regard. This could impact the AI on a more technical and, perhaps, hardware level, as Meta wouldn’t need to force the entirety of Llama 4 to activate just for one task.
Elsewhere, sources claim Meta’s Llama 4 “didn’t perform well” during its internal benchmark tests. Its ability to process math and the AI’s reasoning skills were reportedly the hardest hit areas. As Meta chases the title of best artificial intelligence, another area Llama 4 struggled with was conversations. The company allegedly wants Llama 4 to feel more “humanlike” when engaging with voice — which is exactly what OpenAI did, creating a surge of terrified responses.
Despite all the (reported) internal uncertainty of Llama 4, the publication purports the model is “aiming to release later this month.”
We’ve been expecting Meta’s Llama 4 for a while, ever since the company’s CEO Mark Zuckerbeerg teased its upcoming launch last fall. During its Q3 2024 earnings call, Zuckerberg stated Llama 4 is “expected” to launch in 2025 with improved capabilities. He highlighted that more users have utilized its AI model through Meta AI in places like Facebook, WhatsApp, and its Meta Ray-Ban Smart Glasses. However, the Information cites its alleged sources, stating the company could look to bring Llama 4 to business customers “more effectively.”
Zuckerberg adds Llama 4 is expected to be “much faster” when it (hopefully) launches this year.
Also, Meta announced in February that it plans to bring LlamaCon, a new developer event for those building with Llama, on April 29. The company said it plans to share “the latest on our open source AI developments” to help those developers build apps and products. Meta intends to spend roughly $65 billion on AI this year, so the pressure is definitely on as it continues to develop the next LLM.
With rumors suggesting a Llama 4 launch “this month,” perhaps it’ll line up with LlamaCon — but we’ll have to wait and see.
