OpenAI releases its first open-source AI models in five years

OpenAI has released open-source reasoning AI models. They post strong scores on several benchmarks and are available for download on Hugging Face.

We released two open-weight reasoning models—gpt-oss-120b and gpt-oss-20b—under an Apache 2.0 license.

Developed with open-source community feedback, these models deliver meaningful advancements in both reasoning capabilities & safety.https://t.co/PdKHqDqCPf

— OpenAI (@OpenAI) August 5, 2025

There are two versions:

• the larger, higher-performing gpt-oss-120b — able to run on a single Nvidia GPU;
• a lighter gpt-oss-20b — can run on a laptop with 16 GB of memory.

The release marks OpenAI’s first open-source language models since GPT-2, launched more than five years ago in 2019.

The company noted that its open models can call on more powerful closed LLMs if they cannot solve a task on their own.

In its early years OpenAI published model code, then shifted to a closed development model. That enabled the firm to build a large business by selling access to products via an API.

In January Sam Altman acknowledged the company had been “on the wrong side of history” on open source. It now faces mounting pressure from Chinese labs such as DeepSeek, Alibaba and Moonshot AI, which have built several capable open models.

In July the administration of US president Donald Trump urged American AI developers to disclose technologies more actively to support their global adoption grounded in American values.

Tests, benchmarks, numbers

On the Codeforces programming test, gpt-oss-120b and gpt-oss-20b scored 2622 and 2516 points, respectively. They beat DeepSeek’s R1 but fell short of o3 and o4-mini.

On Humanity Last Exam — a demanding assessment spanning questions across multiple subjects — the models scored 19% and 17.3%, respectively. As in the previous test, they trailed o3 but outperformed leading open models from DeepSeek and Qwen.

In the AIME maths contest the open models scored 96.6 and 96, comparable with the company’s other AIs.

The open networks hallucinate markedly more often than o3 and o4-mini. OpenAI said this is “expected, because small models know less about the world”.

gpt-oss-120b and gpt-oss-20b hallucinated on 49% and 53% of PersonQA questions — an internal benchmark for measuring an AI’s knowledge accuracy about people. O1 has 16%; o4-mini, 36%.

Training method

OpenAI trained the new open models with the same techniques as its closed ones. They are based on a mixture-of-experts architecture, which activates only a subset of parameters for each request. Thus, gpt-oss-120b, with 117bn total parameters, uses 5.1bn per token.

The networks were trained using high-throughput reinforcement learning — a post-training stage in which models acquire behavioural patterns in a simulated environment. Large clusters of Nvidia graphics processors were used.

The same process was used to train the o series. The open models have a similar chain-of-reasoning structure, which requires additional time and compute to work through answers.

OpenAI says the open models suit AI agents: they can call tools such as web search or run Python code while reasoning. However, they are text-only, meaning they cannot process or generate images or audio.

Not entirely transparent

OpenAI released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license, allowing enterprises to monetise OpenAI’s open models without paying or seeking permission from the company.

However, the firm did not publish the training data used to create the networks. The decision is unsurprising, given several lawsuits against AI developers for the “improper” use of information to tune artificial intelligence.

OpenAI delayed the release of open AI models several times over safety concerns. The company examined, among other things, the potential use of gpt-oss for cyberattacks and weapons creation.

The company conducted tests and said the model can slightly increase biological capabilities.

AWS access

With the release of OpenAI’s open models, Amazon said they will be available on Amazon Web Services. This is the first time the startup’s networks are offered on the platform.

Microsoft is also offering versions of the two new AIs optimised for Windows devices.

OpenAI seeks a world-beating valuation

OpenAI is in talks with investors about a share sale. A mooted valuation is $500bn. That would put it ahead of Elon Musk’s SpaceX and make it the most valuable private technology company in the world, writes Financial Times.

At present the firm is closing its most recent funding round at a $300bn valuation. It was reported at the end of March.

In parallel it is already negotiating with Thrive Capital and other investors over a new round. If completed, OpenAI would overtake SpaceX, recently valued at $400bn.

According to the paper, OpenAI is exploring a secondary sale that would allow current and former employees to sell shares. The placement price has not yet been set. The size of the sale will depend on investor demand.

In July OpenAI introduced a special study mode in ChatGPT for students that helps work through tasks step by step, rather than simply receiving a ready-made answer.