Meta AI chief says their coming LLM has caught up with OpenAI's flagship model
摘要
Meta 超人工智能主管 Alexandr Wang 在员工线上会议上透露，下一代模型 Watermelon（继 Avocado 后的模型，目前正在训练中）已追上 OpenAI 的 GPT-5.5，已根据密切跟踪的 AI 模型基准测试得出结论；他补充 Watermelon 的计算量是 Avocado（Muse Spark）的十倍。Wang 同时在 X 平台发帖预告，当前 Muse Spark 即将更新，主要提升代码与 agentic 能力，以缩小与竞争对手模型的差距，回答用户何时有可比 Anthropic Claude Opus 的代码模型时表示「很快」，并称用户即将体验公司正在开发的成果。
荐读理由
水蜜桃模型在训练中，使用比香蕉高一个数量级的算力，已追上GPT-5.5 benchmark，Meta将推更新并瞄准代码与智能体能力赶超Claude Opus
原文
Meta's AI chief, Alexandr Wang. Bloomberg/Getty Images
2026-07-02T23:52:04.159Z
Meta is making significant progress in the AI model race, its superintelligence chief Alexandr Wang told employees today.
In an internal town hall, Wang said that Meta's upcoming AI model — which is codenamed Watermelon — has caught up with OpenAI's flagship GPT-5.5 model, he said, according to two sources familiar with the matter. Wang cited the achievement based on closely followed AI model benchmarks. It's not clear which benchmarks Wang cited.
"Watermelon, our next model after Avocado, is currently in training," Wang said in the town hall, according to a person familiar with the matter. "Watermelon uses an order of magnitude more compute than Avocado," he added, referring to Meta's internal codename for Muse Spark, the first in a family of models that the company released in April.
Wang alluded to that progress publicly, too. In a post on X on Thursday, he said an update to the current model Muse Spark is coming soon, with major gains in coding and agentic capabilities aimed at closing the gap with rival models. Asked by a user when Meta would have a coding model on par with Anthropic's Claude Opus, Wang replied that it would be "pretty soon," adding that users would like what the company has "cooking."
Meta's AI ambitions have long hinged on a simple goal: closing the gap with OpenAI, Google, and Anthropic. Despite a massive investment in chips, data centers, and talent, the company has struggled to convince developers and customers that its models belong at the industry's leading edge.
If Wang's assessment is accurate, it would mark the clearest sign yet that Meta's investment and Zuckerberg's aggressive talent blitz are beginning to pay off, even as the race continues to move at a rapid pace.
GPT 5.5 is a powerful AI model that OpenAI released in April of this year. OpenAI then debuted its most powerful model yet, GPT 5.6, late last month, but hasn't released it generally yet, based on the US government's requests.
Explore BI Games
Take a smarter break in your day - and see how far you get.
Play now
Meta declined to comment. OpenAI didn't respond to a request for comment.
In April, Meta released the first in a series of models called Muse Spark, which performed well on benchmarks but did not match or exceed OpenAI or other labs such as Anthropic.
Zuckerberg is ferociously pushing for Meta to get ahead in the AI race. He appointed Wang last year to head this effort, renaming the company's AI division to Meta Superintelligence Labs.
At Meta, Wang oversees a team of elite AI researchers known as TBD, along with other AI efforts, such as a recent hardware push. Meta has offered top AI talent hundreds of millions of dollars each to join, Business Insider previously reported.
That talent push comes as Meta ramps up spending on infrastructure. The company told investors this year that it expects to spend between $125 billion and $145 billion this year on chips, data centers, and other infrastructure, up from an earlier forecast of $115 billion to $135 billion, citing rising component costs and additional data center spending.
Have a tip? Contact Charles via email at crollet@businessinsider.com or on Signal and WhatsApp at 628-282-2811.* Contact Pranav via email at pdixit@insider.com or on Signal at 408-905-9124. Use a personal email address, a nonwork WiFi network, and a nonwork device; here's our guide to sharing information securely.*
这条对你有帮助吗？