2 weeks ago
9
Fresh from the Ghibli meme wave, OpenAI chief Sam Altman recently confirmed that a powerful new open-weight model with strong reasoning capabilities is on its way. The move is sure to help OpenAI gain more validation and credibility among developers and enterprises, who prefer using open-source weights (not to be confused with open-source models) as it allows them to customise them and host them locally.
“We are planning to release our first open-weight language model since GPT-2. We’ve been thinking about this for a long time, but other priorities took precedence. Now, it feels important to do,” he wrote in his post on X.
This comes in the backdrop of OpenAI raising $40 billion at a $300 billion post-money valuation.
However, why did there happen to be a sudden change of heart? OpenAI doesn’t make decisions without a reason. On the surface, it seems that DeepSeek’s success has influenced this move. Earlier this year, Altman, in a Reddit AMA, said OpenAI has been on the wrong side of history concerning open source.
Previously, Altman even asked users on X whether they prefer an o3-mini-level model or a phone-sized version.
“It’s a welcome move,” said Madhav Krishna, founder of Vahan.ai, in a conversation with AIM about OpenAI’s decision to release open weights. He mentioned that his company will work closely with OpenAI to explore how they can integrate the new model into their operations.
Notably, Vahan.ai uses OpenAI’s API to build AI agents that address hiring challenges for blue-collar workers in India. The platform effectively connects job seekers with appropriate opportunities while assisting employers in identifying the best candidates.
Similarly, another customer of OpenAI APIs, HealthifyMe, told AIM that they are excited to see OpenAI exploring an open-weight model. “While there’s buzz around open-source alternatives like Meta’s Llama and DeepSeek’s R1, for us at Healthify, the main priority has always been robust, reliable personalisation. We’ve built significant value using fine-tuning methods on our existing solutions, which meet our specific health coaching needs,” said Abhijit Khasnis, CTO of HealthifyMe.
“That said, an open-weight model might offer us new ways to tailor outputs for specialised applications or even reduce costs if we can run models on local hardware. However, until we see proven use cases that directly benefit our user engagement or operational efficiencies, we’ll continue to closely monitor the developments,” he added.
Meanwhile, DeepSeek released its advanced reasoning model R1 earlier this year, followed by an upgraded V3 in March, both of which are fully open source. R1 surpassed OpenAI’s o1 model on several benchmarks.
“The biggest revelation from DeepSeek is that open source has won. For a 1% difference in performance, it will be difficult for OpenAI to justify its pricing when the competition is free and formidable,” said Kai-Fu Lee, founder of AI startup 01.AI, in a recent interview with Bloomberg.
He added that the company will also last infinitely because its founder has enough money to fund it at the current level and has reduced computing costs by a factor of five to 10. “With such a formidable competitor, I think Sam Altman is probably not sleeping well,” he quipped.
Why Now?
Open sourcing models will help OpenAI re-engage with the developer community for a more efficient, open model that could lower the cost of its APIs, making them more competitive in the enterprise space.
“OpenAI aims to retain its users by offering an open-source model that can be fine-tuned for specific business needs,” Shantipriya Parida, senior AI scientist at AMD Silo AI, told AIM. He added that with free and better open-source models available, users are questioning the need to spend money on OpenAI subscriptions.
Going by last year’s numbers, OpenAI made $3.4 billion in total revenue, with ChatGPT emerging as the primary revenue driver, contributing $1.9 billion. Meanwhile, revenue from the API was $510 million. This indicates that many developers were not attracted towards OpenAI’s API.
However, according to a recent report by Bloomberg, OpenAI is on track to more than triple its revenue this year, reaching $12.7 billion.
Harneet SN, founder of Rabbitt.ai, told AIM that open-weight models are crucial in sectors like healthcare, defense, and education, where on-premise deployment on their own devices is required. These industries cannot easily delegate or outsource their data to proprietary models.
Having said that, Altman mentioned that the company will now be hosting developer events to gather feedback and later experiment with early prototypes. “We’ll start in San Francisco in a couple of weeks, followed by sessions in Europe and Asia-Pacific.” He further said that he is excited to see what developers build and how large companies and governments use it in cases where they prefer to run a model themselves.
This move will also help OpenAI gain the trust of the US government. Taking a jibe at Meta’s Llama 2 restrictions, Altman said, “We will not do anything silly, like saying that you can’t use our open model if your service has more than 700 million monthly active users. We want everyone to use it!”
Indian IT minister Ashwini Vaishnaw also took a swipe at OpenAI, pointing out how it transitioned its models to closed source after GPT-2, suggesting that Indian open source model makers today might take a similar approach.
“We should also change [OpenAI’s] name,” he quipped.
Not surprisingly, Altman has been buttering up India lately. “What’s happening with AI adoption in India right now is amazing to watch. We love to see the explosion of creativity—India is outpacing the world,” he wrote in a post on X. He even posted a Ghibli-style image of himself playing cricket.
Meanwhile, Meta recently announced that its open-source AI model family, Llama, has surpassed 1 billion downloads. This milestone marks a sharp rise from 650 million downloads in December 2024—a 153% surge in just three months.
Besides Meta, several companies have recently launched open-source models. Mistral’s latest model, Small 3.1, Google’s Gemma 3, and Cohere’s Command A all claim to rival proprietary models while using fewer compute resources.
On the other hand, Chinese tech giants have also upped their game. Alibaba added voice and video chat capabilities to Qwen Chat and released its brand-new open-source model, Qwen2.5-Omni-7B.
Similarly, releasing an open-weight model—where the trained parameters are shared, but the code and training data remain proprietary—will allow OpenAI to stay relevant in a market that increasingly favours accessibility while preserving some control over its intellectual property.
