April 4 (Reuters) - Meta Platforms ( META ) plans to

release the latest version of its large language model later

this month, after delaying it at least twice, the Information

reported on Friday, as the Facebook owner scrambles to lead in

the AI race.

Meta, however, could push back the release of Llama 4 again,

the report said, citing two people familiar with the matter.

Big technology firms have been investing aggressively in AI

infrastructure following the success of OpenAI's ChatGPT, which

altered the tech landscape and drove investment into machine

learning.

The report said one of the reasons for the delay is during

development, Llama 4 did not meet Meta's expectations on

technical benchmarks, particularly in reasoning and math tasks.

The company was also concerned that Llama 4 was less capable

than OpenAI's models in conducting humanlike voice

conversations, the report added.

Meta plans to spend as much as $65 billion this year to

expand its AI infrastructure, amid investor pressure on big tech

firms to show returns on their investments.

Additionally, the rise of the popular, lower-cost model from

Chinese tech firm DeepSeek challenges the belief that developing

the best AI model requires billions of dollars.

The report said Llama 4 is expected to borrow certain

technical aspects from DeepSeek, with at least one version

slated to employ a machine-learning technique called mixture of

experts method, which trains separate parts of models for

specific tasks, making them experts in those areas.

Meta has also considered releasing Llama 4 through Meta AI

first and then as open-source software later, the report said.

Last year, Meta released its mostly free Llama 3 AI model,

which can converse in eight languages, write higher-quality

computer code and solve more complex math problems than previous

versions.