OpenAI introduced Tuesday the launch of two open-weight AI reasoning fashions with related capabilities to its o-series. Each are freely accessible to obtain from the net developer platform Hugging Face, the corporate stated, describing the fashions as “state of the art” when measured throughout a number of benchmarks for evaluating open fashions.
The fashions are available in two sizes: a bigger and extra succesful gpt-oss-120b mannequin that may run on a single Nvidia GPU, and a lighter-weight gpt-oss-20b mannequin that may run on a shopper laptop computer with 16GB of reminiscence.
The launch marks OpenAI’s first ‘open’ language mannequin since GPT-2, which was launched greater than 5 years in the past.
In a briefing, OpenAI stated its open fashions shall be able to sending complicated queries to AI fashions within the cloud, as TechCrunch beforehand reported. Which means if OpenAI’s open mannequin shouldn’t be able to a sure process, similar to processing a picture, builders can join the open mannequin to one of many firm’s extra succesful closed fashions.
Whereas OpenAI open sourced AI fashions in its early days, the corporate has typically favored a proprietary, closed supply improvement strategy. The latter technique has helped OpenAI construct a big enterprise promoting entry to its AI fashions by way of an API to enterprises and builders.
Nonetheless, CEO Sam Altman stated in January he believes OpenAI has been “on the wrong side of history” relating to open sourcing its applied sciences. The corporate at this time faces rising strain from Chinese language AI labs — together with DeepSeek, Alibaba’s Qwen, and Moonshot AI — which have developed a number of of the world’s most succesful and widespread open fashions. (Whereas Meta beforehand dominated the open AI house, the corporate’s Llama AI fashions have fallen behind within the final 12 months.)
In July, the Trump administration additionally urged U.S. AI builders to open supply extra know-how to advertise world adoption of AI aligned with American values.
Techcrunch occasion
San Francisco
|
October 27-29, 2025
With the discharge of gpt-oss, OpenAI hopes to curry favor with builders and the Trump administration alike, each of which have watched the Chinese language AI labs rise to prominence within the open supply house.
“Going back to when we started in 2015, OpenAI’s mission is to ensure AGI that benefits all of humanity,” stated Altman in an announcement shared with TechCrunch. “To that end, we are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.”
How the fashions carried out
OpenAI aimed to make its open mannequin a frontrunner amongst different open-weight AI fashions, and the corporate claims to have achieved simply that.
On Codeforces (with instruments), a aggressive coding take a look at, gpt-oss-120b and gpt-oss-20b rating 2622 and 2516, respectively, outperforming DeepSeek’s R1 whereas underperforming o3 and o4-mini.
On Humanity’s Final Examination (HLE), a difficult take a look at of crowdsourced questions throughout quite a lot of topics (with instruments), gpt-oss-120b and gpt-oss-20b rating 19% and 17.3%, respectively. Equally, this underperforms o3 however outperforms main open fashions from DeepSeek and Qwen.
Notably, OpenAI’s open fashions hallucinate considerably greater than its newest AI reasoning fashions, o3 and o4-mini.
Hallucinations have been getting extra extreme in OpenAI’s newest AI reasoning fashions, and the corporate beforehand stated it doesn’t fairly perceive why. In a white paper, OpenAI says that is “expected, as smaller models have less world knowledge than larger frontier models and tend to hallucinate more.”
OpenAI discovered that gpt-oss-120b and gpt-oss-20b hallucinated in response to 49% and 53% of questions on PersonQA, the corporate’s in-house benchmark for measuring the accuracy of a mannequin’s data about individuals. That’s greater than triple the hallucination charge of OpenAI’s o1 mannequin, which scored 16%, and better than its o4-mini mannequin, which scored 36%.
Coaching the brand new fashions
OpenAI says its open fashions have been skilled with related processes to its proprietary fashions. The corporate says every open mannequin leverages mixture-of-experts (MoE) to faucet fewer parameters for any given query, making it run extra effectively. For gpt-oss-120b, which has 117 billion whole parameters, OpenAI says the mannequin solely prompts 5.1 billion parameters per token.
The corporate additionally says its open mannequin was skilled utilizing high-compute reinforcement studying (RL) — a post-training course of to show AI fashions proper from improper in simulated environments utilizing massive clusters of Nvidia GPUs. This was additionally used to coach OpenAI’s o-series of fashions, and the open fashions have an analogous chain-of-thought course of during which they take further time and computational sources to work by their solutions.
On account of the post-training course of, OpenAI says its open AI fashions excel at powering AI brokers and are able to calling instruments similar to net search or Python code execution as a part of its chain-of-thought course of. Nonetheless, OpenAI says its open fashions are text-only, which means they will be unable to course of or generate photographs and audio like the corporate’s different fashions.
OpenAI is releasing gpt-oss-120b and gpt-oss-20b below the Apache 2.0 license, which is usually thought of one of the vital permissive. This license will permit enterprises to monetize OpenAI’s open fashions with out having to pay or receive permission from the corporate.
Nonetheless, in contrast to absolutely open supply choices from AI labs like AI2, OpenAI says it won’t be releasing the coaching information used to create its open fashions. This determination isn’t a surprise on condition that a number of energetic lawsuits towards AI mannequin suppliers, together with OpenAI, have alleged that these firms inappropriately skilled their AI fashions on copyrighted works.
OpenAI delayed the discharge of its open fashions a number of occasions in latest months, partially to deal with security issues. Past the corporate’s typical security insurance policies, OpenAI says in a white paper that it additionally investigated whether or not unhealthy actors may fine-tune its gpt-oss fashions to be extra useful in cyberattacks or the creation of organic or chemical weapons.
After testing from OpenAI and third-party evaluators, the corporate says gpt-oss might marginally enhance organic capabilities. Nonetheless, it didn’t discover proof that these open fashions may attain its “high capability” threshold for hazard in these domains, even after fine-tuning.
Whereas OpenAI’s mannequin seems to be state-of-the-art amongst open fashions, builders are eagerly awaiting the discharge of DeepSeek R2, its subsequent AI reasoning mannequin, in addition to a brand new open mannequin from Meta’s Superintelligence Lab.
