Meta Platforms is set to unveil the next iteration of its large language model later this month, following multiple delays, according to a report by the Information released on Friday. The company is striving to maintain its position at the forefront of the AI sector.
However, there remains a possibility that the launch of Llama 4 could be postponed once more, as indicated by sources familiar with the situation.
Major tech companies have been aggressively increasing their investments in AI infrastructure in the wake of OpenAI’s ChatGPT, which significantly impacted the technology sector and spurred heightened interest in machine learning.
The report highlights that one of the factors behind the delays is that Llama 4 has not met Meta’s expectations in certain technical benchmarks, especially concerning reasoning and mathematical abilities.
Concerns were also raised regarding Llama 4’s performance in generating humanlike voice interactions, which appeared to fall short when compared to OpenAI’s models, the report noted.
This year, Meta plans to allocate up to $65 billion (approximately Rs. 5,39,000 crore) towards enhancing its AI infrastructure, responding to investor demands for visible returns on their investments in the technology sector.
Furthermore, the emergence of a competitively priced model from Chinese firm DeepSeek challenges the notion that creating an advanced AI model necessitates substantial funding.
According to the report, Llama 4 is expected to integrate certain technical features from DeepSeek’s offerings, with at least one version anticipated to utilize a technique known as the mixture of experts method. This approach enables different model components to specialize in specific tasks, enhancing their effectiveness in those areas.
Meta is also exploring the option of initially launching Llama 4 through Meta AI before eventually releasing it as open-source software, the report states.
Last year, the company introduced the largely free Llama 3 AI model, which boasts the capability to converse in eight languages, produce higher-quality computer code, and tackle more intricate mathematical problems compared to its predecessors.
© Thomson Reuters 2025
