OpenAI takes an important step towards open-source models: launching open-weight models for the first time in six years.
On Tuesday, August 5th, Eastern Time, OpenAI released two free-to-use open-weight language models, gpt-oss-120b and gpt-oss-20b. This is the first release of open-weight models by OpenAI since the launch of GPT-2 in 2019, and these are the first such models to emerge since OpenAI signed an exclusive cloud service agreement with Microsoft six years ago.
The two new models can generate text content based on user prompts and allow developers to customize them, but OpenAI has not provided the data used to train them. Commentators believe that OpenAI's new models are a direct response to challenges from competitors in the open-weight model space, such as Meta, Microsoft-backed Mistral AI, and China's startup DeepSeek.
OpenAI's CEO Sam Altman stated on social media:
<blockquote>
&#34;gpt-oss is a significant breakthrough; it is the most advanced open-weight inference model with powerful real-world performance comparable to o4-mini, capable of running locally on your own computer (or a smaller version on your phone). We believe this is the best and most practical open model in the world.&#34;
</blockquote>
gpt-oss-120b and gpt-oss-20b are designed as low-cost GPT model options that developers, researchers, and businesses can easily run and customize. gpt-oss-20b can run on laptops with 16GB of memory, while gpt-oss-120b requires about 80GB of memory. Both models support advanced reasoning, tool usage, and chain-of-thought processing.
Altman noted that these two models can run locally on personal computers (PCs) and even phones without an internet connection, providing users with complete control and privacy protection.
At the same time, Amazon announced that it will offer OpenAI's models to customers for the first time, planning to provide OpenAI's open AI weight models on its Bedrock and SageMaker platforms. This marks the first time the cloud computing giant Amazon has offered OpenAI's products.
The release of gpt-oss-20b and 120b marks an important turning point in OpenAI's strategy, as the company had previously delayed the release of open-weight models multiple times and conducted extensive safety testing and evaluations to ensure the models are not maliciously exploited.
The so-called open weights represent a middle ground between open-source and closed-source, meaning that businesses and government agencies can run the model independently, as they can view and modify the model's weights. In the case of open-source, users can view the source code composition of the model, sometimes including the training and weight distribution methods.
Previously, it was suggested that OpenAI's launch of open-weight models could weaken Microsoft's exclusive advantage in the AI business. On Tuesday, OpenAI stated that Microsoft would introduce a GPU-optimized version of the gpt-oss-20b model for Windows devices. These models are supported by ONNX Runtime, enabling local inference and can be accessed through Foundry Local and AI Toolkit for VS Code, making it easier for Windows developers to build using open models
<h2>The performance of gpt-oss-120b in some tests even surpasses that of o4-mini, while 20b surpasses o3-mini in some tests</h2>
OpenAI introduced that the two gpt-oss models utilize its most advanced pre-training and post-training techniques, with a particular focus on reasoning, efficiency, and practicality across deployment environments. Both models are based on a Transformer architecture using a mixture of experts (MoE) to reduce the number of active parameters required for processing inputs.
gpt-oss-120b activates 510 million parameters per token, with a total parameter count of 117 billion; gpt-oss-20b activates 3.6 billion parameters, with a total parameter count of 21 billion. The models use alternating dense and locally banded sparse attention patterns, supporting context lengths of up to 128k.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/ab086b4d-bd33-4afe-8854-277bfef02bdc.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/ab086b4d-bd33-4afe-8854-277bfef02bdc.jpeg"/>
OpenAI describes gpt-oss-20b as a medium-sized open model, while gpt-oss-120b is a large open model that can run on &#34;most desktops and laptops.&#34;
OpenAI states that the smaller model requires at least 16GB of video memory or unified memory to perform optimally and is &#34;very suitable for high-end consumer GPUs and Mac computers with Apple chips.&#34; The larger full-size model requires at least 60GB of video memory or unified memory.
In terms of performance evaluation, OpenAI disclosed that gpt-oss-120b performs close to OpenAI o4-mini in core reasoning benchmark tests. The smaller gpt-oss-20b model achieves or exceeds the performance of OpenAI o3-mini in the same evaluations.
gpt-oss-120b outperforms OpenAI o3-mini in competitive coding (Codeforces), general problem solving (MMLU and HLE), and tool invocation (TauBench), and it matches or even surpasses OpenAI o4-mini. Its performance in health-related queries (HealthBench) and competitive mathematics (AIME 2024 and 2025) even exceeds that of o4-mini.
Even at a smaller scale, gpt-oss-20b matches or even surpasses OpenAI o3-mini in the aforementioned evaluations, even exceeding the latter in competitive mathematics and health.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/9f6bf4ce-6c59-4359-a790-6eb1448c9f2c.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/9f6bf4ce-6c59-4359-a790-6eb1448c9f2c.jpeg"/>
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2bd18abf-fa92-4880-825c-d2c39f68e436.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2bd18abf-fa92-4880-825c-d2c39f68e436.jpeg"/>
<h2>Security Protection and Prevention of Malicious Use</h2>
OpenAI has paid special attention to the security of open-weight models, as once released, adversaries may maliciously fine-tune the models. The company filtered harmful data related to chemical, biological, radiological, and nuclear (CBRN) during the pre-training phase.
To directly assess risks, OpenAI fine-tuned the model with specialized biological and cybersecurity data, creating specific domain non-reject versions that attackers might produce. After internal and external testing, these maliciously fine-tuned models did not reach the &#34;high capability&#34; threshold in their preparedness framework.
OpenAI also collaborated with three independent expert groups to provide feedback on its malicious fine-tuning assessments. Altman emphasized, &#34;We strive to mitigate the most severe security issues, particularly in biosafety. The gpt-oss model performed comparably to our cutting-edge models in internal security benchmark tests.&#34;
<h2>Ecosystem Deployment and Market Impact</h2>
The two models are available for free under the Apache 2.0 license, and users can download them from platforms like Hugging Face and GitHub. Cloud service providers Amazon, Baseten, and Microsoft will also offer these models.
OpenAI has partnered with leading deployment platforms including Azure, vLLM, Ollama, AWS, and Fireworks to enable developers to widely use these models. Microsoft will also introduce a GPU-optimized version of the gpt-oss-20b model for Windows devices.
The company is hosting a red team challenge with a prize pool of $500,000, encouraging global researchers, developers, and enthusiasts to help identify new security issues. Altman stated, &#34;We believe this will bring more benefits than harms, and we look forward to this release enabling new types of research and creating new types of products.&#34;
<h2>Competitive Pressure Driving Strategic Shift</h2>
OpenAI's release of open-weight models is largely driven by competitive pressure in the market, particularly from the challenges posed by China's DeepSeek. Since 2019, the company has focused on providing proprietary model services through APIs, while competitors have launched open-weight alternatives.
OpenAI President Greg Brockman stated at a press conference on Tuesday, &#34;It's exciting to see the development of the ecosystem, and we are pleased to contribute to it, truly pushing the frontier and then seeing what happens.&#34;
OpenAI has collaborated with chip manufacturers such as Nvidia, AMD, Cerebras, and Groq to ensure the models run well on various chips. Nvidia CEO Jensen Huang stated in a statement, &#34;OpenAI has shown the world what can be built on Nvidia AI—now they are advancing innovation in open-source software.&#34;

AMZN

OpenAI

Altman stated that gpt-oss is the most advanced open-weight inference model. OpenAI claimed that gpt-oss-120b performs close to o4-mini in core inference benchmark tests, and even surpasses it in health-related queries and competitive mathematics; the smaller gpt-oss-20b model achieved or exceeded o3-mini performance in evaluations; the smaller models are suitable for Mac computers equipped with Apple chips. Amazon announced that its platforms Bedrock and SageMaker will offer these two models, marking the company's first provision of OpenAI models to customers

- OpenAI has released two open-weight language models, gpt-oss-120b and gpt-oss-20b, marking its first such release since 2019.  
- These models allow developers to generate text based on prompts and can run locally on personal computers, enhancing user control and privacy.  
- The release responds to competitive pressures in the open-weight model space and is part of OpenAI's strategic shift to address challenges from rivals like Meta and DeepSeek.

For the first time in six years! OpenAI's new model opens weights, Altman calls it "the world's best open model"