Meta's next-generation AI model Avocado has completed pre-training, internally referred to as the "most powerful" foundational model
According to The Information, Megan Fu, product manager of Meta's Super Intelligence Lab, disclosed in an internal memo that Avocado has completed its pre-training phase, outperforming leading open-source pre-trained foundational models in testing. Although it has not yet undergone subsequent training optimization, the model is already able to compete with leading post-trained models in knowledge, visual perception, and multilingual performance
Meta informed some employees that the internally codenamed Avocado model has become the company's "most powerful pre-trained foundational model to date."
On February 4th, The Information reported that Megan Fu, product manager at Meta's Superintelligence Lab, disclosed new progress on the Avocado model in an internal memo dated January 20th.
The memo indicates that Avocado has completed the pre-training phase, which is the initial stage of AI model development where the model learns general knowledge, patterns, and relationships by accessing large datasets.
The memo states that Avocado has outperformed leading open-source pre-trained foundational models in testing. Although it has not yet undergone subsequent training optimization, the model is already able to "compete" with leading post-trained models in terms of knowledge, visual perception, and multilingual performance.
This progress is significant for Meta's competitive position in the AI field, but the actual performance of the model remains to be independently verified after its public release. On Wednesday, Meta's stock price fell by 3.28%.
Superintelligence Lab Leads Development
The Avocado model is developed by Meta's Superintelligence Lab, which is the internal team specifically responsible for AI research and development. This lab represents the latest organizational restructuring of Meta in the field of artificial intelligence.
Pre-training is a critical initial phase in AI model development, during which the model acquires general knowledge and patterns by accessing large-scale datasets before being fine-tuned for specific tasks. Avocado has currently completed this phase.
According to the internal memo, Avocado's competitiveness across multiple dimensions has been achieved without undergoing post-training optimization. Post-training is the next phase of model development, typically used for tuning and improving the model for specific application scenarios.
Although there is optimism within Meta regarding Avocado's capabilities, the actual performance of the model cannot currently be independently verified. Only when the model is publicly released can external researchers and developers objectively assess its capabilities.
Reportedly, Meta has not disclosed a timeline for the release of the Avocado model or its commercialization plans. The company has also not revealed whether the model will be released in an open-source manner or be for internal use only
