OpenAI launches free, fully customizable AI models to directly challenge Meta and DeepSeek’s offerings

OpenAI launches free, fully customizable AI models to directly challenge Meta and DeepSeek’s offerings

OpenAI has unveiled two new “open-weight” language models, making its own AI technology freely downloadable and customizable in a direct challenge to Mark Zuckerberg’s Meta and China’s DeepSeek.

The developer behind ChatGPT announced the release of gpt-oss-120b and gpt-oss-20b-two, each designed for public download and fine-tuning. This move departs from OpenAI’s flagship ChatGPT, which runs on a proprietary model that users cannot alter. Sam Altman, OpenAI’s chief executive, said the company is proud to add its work—“the fruit of billions of dollars in research”—to a growing library of AI grounded in democratic values and accessible to all.

Altman noted that the new models are built to support autonomous AI agents and fit seamlessly into complex “agent workflows.” Zuckerberg echoed this sentiment, arguing that open access and customization ensure AI’s benefits reach a broader audience rather than concentrating power among a few corporations.

Meta, however, has cautioned that it may need to exercise restraint before releasing equally advanced open models. DeepSeek has already offered comparable free, customizable systems. In head-to-head tests, OpenAI reports that its largest open-weight model matches the reasoning performance of its o4-mini model and outstrips other models of similar size.

To gauge potential misuse, OpenAI created “maliciously modified” variants of its models, simulating scenarios ranging from biothreat creation to cyberattacks. The company found these doctored versions fell short of posing serious risks—but experts warn that any powerful, freely available AI could be repurposed for harmful ends.

Meta brands its Llama line as open source—providing full datasets, architecture, and training code—yet the Open Source Initiative has ruled that Meta’s licensing constraints keep it below true open-source status. OpenAI sidesteps that debate by calling its offering “open weight,” a designation that allows fine-tuning without granting complete operational transparency.

OpenAI’s announcement arrives amid widespread speculation that a new ChatGPT backbone model—potentially GPT-5—is imminent; Altman shared what appeared to be an early GPT-5 interface over the weekend. At the same time, Google is pushing its own frontier toward artificial general intelligence with Genie 3, a “world model” that lets AI agents and robots train inside realistic simulations of environments such as warehouses.

DeepMind, Google’s AI arm, describes world models as essential steps toward AGI—a level of machine intelligence capable of performing a broad spectrum of tasks on par with humans. “We expect this technology to play a pivotal role as we move toward AI and as agents play a larger role in the world,” DeepMind said.