Meta Developing Open-Source Version of Upcoming AI Model

Meta Platforms Inc. is preparing to launch an open-source variant of its next-generation artificial intelligence model.

The companyinitially launchedits most powerful neural network last April. The algorithm, known as Llama 4 Maverick, is an open-source large language model with 400 billion parameters. Last December, sources toldBloombergMeta plans to shift to closed-source distribution for future LLMs.

The company isreportedlydeveloping two proprietary frontier models. One is an LLM codenamed Avocado, and the other is a multimedia file generator internally called Mango. Both algorithms are expected to launch this year. The open-source model detailed in today's Axios report may derive from Avocado and Mango.

It remains unclear if the open-source version will debut simultaneously with the proprietary one. Axios reports that the former algorithm will be released "eventually." This is part of Meta's effort to distribute its models widely across the globe.

Reportedly, the open-source iteration will not include all features available in the closed-source version. It is currently unclear which functionalities Meta will omit.

Llama 4 Maverick stands as the company's premier open-source LLM, built on a Mixture of Experts architecture. It is not a single algorithm but comprises 128 distinct neural networks, each optimized for specific tasks. Meta's upcoming open-source model might lack some neural networks supporting the proprietary version.

Another possibility is that the company will reduce the parameter count for the open-source version or skip certain training steps. Frontier LLMs undergo multiple training rounds, including so-called post-training after core capabilities are established.It is not uncommon for open-source model developers to release reduced versions without post-training.AI safety is reportedly one reason Meta's open-source model will not contain all proprietary features. This suggests Avocado will excel at generating cybersecurity-related code. Claude 4.6 Opus, Anthropic PBC's most powerful LLM, has so far

identifiedhundreds of critical vulnerabilities in open-source projects.

Both Anthropic and OpenAI Group PBC arepreparingto release new flagship LLMs. According to Axios, Meta does not expect its upcoming model to surpass competitors "comprehensively." However, the Facebook parent company's algorithm reportedly has potential to attract consumers in multiple "strength areas."

Meta may attract consumers by making its models more hardware-efficient than competitors. Many frontier LLMs cannot run on personal computers due to processor limitations. Meta might also optimize algorithms for use cases like personal health and homework assistance, which enterprises do not prioritize.