Openai has announced plans to release its first open-weight language model with inference capabilities since GPT-2. CEO Sam Altman shared the update Monday, saying the model could be available in the coming months.
This shows a major change in Openai’s direction. After years of maintaining the most advanced models behind closed APIs and subscriptions, the company is opening doors again. With inference built-in, this new model doesn’t just generate text. It is intended to think through problems, create logical connections, and provide a more structured response.
Altman said, like other releases, the model goes through an internal OpenAI evaluation process before launching. But this time they’ve added more checks and know well enough that if the model goes out in the wild it changes the model.
“TL;DR: We look forward to releasing a powerful new openweight language model with inference in the coming months, and we’d like to talk to developers about how to make it the most useful,” Altman told X.
This is a pretty big moment for Openai. Since the release of the GPT-2 in 2019, the company has been firmly covering the new models, offering access only via APIs or paid subscriptions. This time, things are different. Future models can be downloaded and changed. This is a shift that can ripple over AI space.
“We’ve been thinking about this for a long time, but other priorities have been prioritized,” writes Altman. “I feel it’s important now,” he adds that the model will run through Openai’s prep framework before launch, and he acknowledges that the team is doing additional prep work, and knows well enough that developers will change it when they go out.
Openai starts things with a developer event in San Francisco, followed by sessions in Europe and Asia Pacific. The goal is to gather feedback, test early prototypes, and better understand how developers can use the model before they officially launch.
“I’m looking forward to what developers will build and how they will use it where big companies and governments prefer to run their models on their own,” added Altman.
What is an open weight model?
The open weight model allows developers to access internal parameters directly in neural networks. Basically, it is the “gut” that works. Unlike cloud-based models that query with APIs, these models can be downloaded and run locally or on a custom infrastructure. As a result, they are more flexible, cheaper to use at scale, and easier to customize for specific use cases.
Think of it as having access to source code instead of locked apps. It’s not just using models. You can change it as you like.
Why now?
It’s not just the nostalgia of GPT-2 that drives this decision. The market has changed. Meta started things with the Llama model in 2023, and since then other players like Deepseek have dived into their own openweight alternatives. Deepseek’s R1 model has particularly been noted that it offers strong performance at a small fraction of the costs of larger Western models.
Altman is not shy about accepting shifts. I feel the timing is right now. Competitive pressure could also play a role here. Openai has created a name for a premium AI product, but it’s hard to ignore the momentum around open source alternatives.
Releasing an open weight model is not just a move to maintain its relevance. This is a step towards regaining influence in part of the AI ecosystem that helped companies once define.
What developers can expect
Openai invites developers to be part of the process. Through upcoming events and dedicated early access pages, the company is trying to understand what users actually need and build it.
This approach can make a real difference. By allowing developers to run their models locally, Openai offers the opportunity to build on small teams and organizations without resorting to expensive third-party tools. The possibilities are widespread, from startups to universities and public institutions.
And for developers, there’s more to more than cost savings. Openweight models can be tweaked for specific tasks and integrated into custom workflows. This is difficult to do with the Lock-In Cloud API. Gain freedom and control. Developers tend to value it.
Some risks of mixing
Still, it’s not all a smooth voyage. When such a model comes out in the wild, the control becomes even more tricky. Openai deals with a significant proportion of scrutiny: copyright litigation, content moderation challenges, and concerns about misuse. A model that is superior to previous versions may raise new questions about safety and accountability.
There is also competition. For example, Deepseek’s R1 is already known for its speed and cost-effectiveness. If Openai wants developers to take this release seriously, it needs to prove that it is more than just powerful, it is worth the switch.
Looking ahead
As of March 31, 2025, all eyes are in Openai. This upcoming model could mark a shift in how the company engages with the broader tech community and how developers build with AI.
The interests are high, but the excitement is high. Much depends on execution, transparency, and whether Openai can balance openness and responsibility.
Altman said it best: “Now I feel it’s important.” If they pull it off, this release could be one of the most influential things Openai has done in years.
Summer isn’t that far. If this lands as expected, it can be a big deal.
Source link
