The technological firm Openai He has launched his new O3-mini reasoning model, which is the “most profitable” he has announced to date, has an “exceptional” performance in Stem disciplines and incorporates reasoning capabilities for the first time for free users.
O3-mini offers performance comparable to Openai O1 in mathematics, coding and science, But with a “significantly lower” price and faster response times, as explained by the company in a statement.
The technology company stressed that this new model It is the first of small size that admits developer functions “very requested”including the call of functions -which allows the models to obtain data and perform actions -indications for developers or structured outputs, attached to a JSON scheme. Likewise, as OpenAi O1-mini and OpenAi O1 Preview, he will admit ‘streaming’.
That model, which is “The most profitable” Because it has also reached users freecan be adjusted to different levels of reasoning effort (with low, medium and high options) to optimize specific use cases, thanks to which it can “think more” by addressing complex challenges.
Instead, O3-mini does not admit vision capabilitiesas regards developers to use OPENAI O1 to execute visual reasoning tasks. It also responds 24 percent faster than O1-mini, maintaining exceptional performance in Stem (science, technology, engineering and mathematics).
In comparison to this other, external evaluators prefer the responses of the new model by 56 % more of the time and detect 39 % less significant errors in complex questions of the real world. Likewise, the firm has advanced that O3-mini exceeds OpenAi O1 in various coding and medium level reasoning tasks.
On the other hand, the company has indicated that one of the key techniques they have used to teach the model to respond safely is Deliberative alignmentin which they have trained the model so that RAzone about the security specifications written by humans before responding to the user’s instructions.
OpenAI O3-MINI, optimized for production environments, is already being implemented in ChatgPT and in the application programming interfaces (API) of completion of conversations, assistants and lots.