OpenAI Launches Two Powerful Open-Weight AI Models, Marking Its Return to Open Source After 5 Years

August 6, 2025 — In a landmark shift for the AI ecosystem, OpenAI has launched two open-weight AI reasoning models, marking the company’s first open-source model release since GPT-2 in 2019. The two models, gpt-oss-120b and gpt-oss-20b, are now available for free download on the popular developer platform Hugging Face under the permissive Apache 2.0 license, allowing developers and enterprises to use, fine-tune, and even commercialize them.

According to OpenAI, the new models are “state of the art” among publicly available alternatives and demonstrate significant reasoning capabilities while remaining highly efficient and lightweight compared to their closed counterparts.

📦 What’s in the Release?

The two models come in different sizes to support a wide range of developers:

• gpt-oss-120b: A larger model designed to run on a single Nvidia A100 or similar high-end GPU.
• gpt-oss-20b: A smaller variant that can run on a consumer-grade laptop with 16GB RAM.

Both models were trained using mixture-of-experts (MoE) architecture, allowing only a fraction of the model’s parameters to activate during inference. This significantly reduces computational demands: for example, gpt-oss-120b uses just 5.1 billion active parameters per token despite having 117 billion total.

🧠 Advanced Reasoning with Limitations

Though the models are open, they retain advanced reasoning features, including support for tool use—such as invoking a web search or Python function call during complex queries. The models are text-only, however, meaning they cannot handle images or audio inputs like OpenAI’s proprietary multimodal models (e.g., GPT-4o).

One of the models’ standout features is the ability to offload more complex requests to OpenAI’s more advanced closed models. For instance, if gpt-oss can’t handle a visual task, it can be designed to forward that task to GPT-4 or GPT-4o in the cloud, enabling hybrid use cases.

📊 How Do They Perform?

According to OpenAI, gpt-oss models outperform most leading open-weight models in benchmark tests, though they still lag behind OpenAI’s proprietary offerings like o3 and o4-mini.

• On Codeforces (with tools):
  • gpt-oss-120b: 2622
  • gpt-oss-20b: 2516
  • (Beating DeepSeek R1, trailing o-series models)
• On Humanity’s Last Exam (HLE):
  • gpt-oss-120b: 19%
  • gpt-oss-20b: 17.3%
  • (Again better than DeepSeek and Qwen but behind o3)

However, hallucination rates remain high—with gpt-oss-120b hallucinating on 49% of PersonQA prompts, and gpt-oss-20b at 53%, far above the 16% of o1 and even higher than o4-mini’s 36%. OpenAI attributes this to the smaller models’ limited “world knowledge.”

🔐 Open, But Not Fully Transparent

While the models are freely available under Apache 2.0, OpenAI has chosen not to release the training data. This is a key distinction from labs like Allen Institute (AI2) or Meta’s open-source efforts. The decision is likely linked to ongoing lawsuits accusing AI firms of training on copyrighted content without consent.

OpenAI CEO Sam Altman acknowledged the shift, saying in a statement:

“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.”

The move also follows pressure from the Trump administration, which in July called for U.S. AI companies to open-source more technology in order to counter China’s growing dominance in AI development. Chinese labs like DeepSeek, Alibaba’s Qwen, and Moonshot AI have recently produced some of the most advanced open models globally, challenging U.S. leadership in this space.

🧪 Safety Considerations and Delays

OpenAI delayed the release of the gpt-oss models multiple times, citing safety concerns. In a technical white paper released alongside the models, OpenAI says it evaluated how easily these open models could be fine-tuned to aid cyberattacks or bioweapon creation.

While the models marginally increased biological knowledge, the company and third-party testers did not find them capable of reaching high-risk thresholds, even with fine-tuning.

This extra layer of vetting reflects OpenAI’s efforts to stay ahead of criticism around unsafe model deployment—a challenge for all AI developers working with open-weight architectures.

💼 Enterprise-Ready, Developer-Friendly

Releasing the models under Apache 2.0 makes them attractive to businesses. No licensing fees or usage restrictions are required, making it easy for companies to build proprietary applications on top of OpenAI’s base model.

Use cases include:

• AI agents for customer service or technical support
• Code generation tools
• Educational tutors
• Data analysis chatbots

In particular, the gpt-oss-20b is expected to gain traction among startups and independent developers, as it can be fine-tuned and run on standard hardware.

🔭 Looking Ahead: The Global Open AI Race

While OpenAI’s release is a milestone, the competition is fierce. Developers are awaiting:

• DeepSeek R2 (expected to be significantly more powerful)
• Meta Superintelligence Lab’s next-gen open model
• Updates from Mistral, AI2, and Cohere

Still, OpenAI’s return to open models—backed by sophisticated tooling, safety protocols, and cloud connectivity—may reshape how developers view the company in the long run.

By blending open access with enterprise-grade features, OpenAI is positioning gpt-oss not only as a research contribution but as a strategic counterbalance to global competitors and a new foundation for innovation at scale.

🧾 Quick Facts

With this release, OpenAI reclaims a seat at the open-source table, setting a new standard for AI accessibility and signaling a broader industry shift toward openness with responsibility.