[AI Landscape Shift] OpenAI Embraces Openness After Six Years, While Meta Steps on the Brakes

The AI world is witnessing a surprising shift in attitudes. Once heavily criticized for its closed-source stance, OpenAI has finally taken a major step toward openness after six years. Meanwhile, Meta, once a leading advocate of open-source AI, is now calling for “cautious openness.” The future of open AI ecosystems is entering a new phase of divergence.

🧠 OpenAI Returns to “Open” with Two Released Model Weights

On August 5 (local time), OpenAI released two open-weight models: gpt-oss-120b and gpt-oss-20b, now available for free download on Hugging Face. These are the first public model weights OpenAI has offered since GPT-2 in 2019—a significant strategic pivot.

OpenAI claims that both models outperform other open models in their respective classes and are optimized for consumer-grade hardware:

• gpt-oss-120b performs comparably to o4-mini and runs on a single 80GB GPU;
• gpt-oss-20b is on par with o3-mini and runs on laptops with 16GB RAM.

It’s important to note that these models are open-weight but not fully open-source—OpenAI disclosed parameters but withheld training data and procedures. This “partial openness” approach is now widely adopted across the AI field, as seen in Meta’s LLaMA, Alibaba’s Qwen, and DeepSeek.

OpenAI stated that the goal is to reduce barriers for developers in emerging markets, small teams with limited resources, or those requiring budget flexibility.

“A healthy open model ecosystem is crucial to making AI broadly accessible and beneficial to all,” OpenAI wrote.

🤖 From "ClosedAI" to Public Reflection

OpenAI's prior closed strategy often drew criticism, including from Elon Musk, who once joked the company should be renamed "ClosedAI." But with the rise of open models like DeepSeek, even OpenAI CEO Sam Altman admitted,

“We might have been on the wrong side of history with our stance.”

And OpenAI isn't alone in pivoting. Earlier this year, Baidu’s CEO Robin Li acknowledged during an earnings call that DeepSeek taught them an important lesson:

“The best open-source models can drive organic adoption.”
In June, Baidu open-sourced its Wenxin 4.5 models, reversing its previous stance.

⚠️ Meta’s Changing Tune: From Open to Cautious?

Surprisingly, while the open-source movement gains traction, Meta has become more cautious. On July 30, Meta CEO Mark Zuckerberg stated in an open letter that

“superintelligence poses real risks, and we must be selective about what we open-source.”

Just a year ago, he had declared that “open-source is the path forward,” predicting that LLaMA would lead the AI race by 2025. But Meta has recently stumbled:

• The LLaMA 4 series fell short of performance expectations;
• DeepSeek’s success applied competitive pressure.

In response, Zuckerberg launched an aggressive strategy:

• Invested $14.3 billion in Scale AI, a major data-labeling company, and recruited its CEO;
• Poured billions into hiring top AI talent from OpenAI, Google, and others;
• Launched Meta Superintelligence Labs, aiming to develop universally accessible personal superintelligence.

Meta insists its stance hasn't changed:

“We will continue to release leading open models,” a spokesperson said.
“Our development roadmap includes both open and closed models.”

✅ Conclusion: A Fragmented Future for Open AI

OpenAI is cautiously stepping back into openness, while Meta is shifting into selective transparency. In an era dominated by large language models and AI proliferation, the debate around open vs closed AI is no longer purely technical—it’s strategic, ethical, and geopolitical.

Whether the future will favor fully open ecosystems or a hybrid landscape remains to be seen—but the battle lines are clearly being drawn.