OpenAI Breakthrough Release of Two Open-Source Reasoning Models gpt-oss-120b and gpt-oss-20b: First Time in Five Years to Open Core Technology

August 06, 2025
OpenAI
5 min

News Summary

OpenAI Makes Breakthrough Release of Two Open-Source Inference Models, gpt-oss-120b and gpt-oss-20b, Opening Core Technology for the First Time in Five Years

Major Strategic Shift: OpenAI Returns to Open-Source Path

August 6, 2025 – The artificial intelligence sector witnessed a significant announcement today. OpenAI officially released two open-source weight inference models, gpt-oss-120b and gpt-oss-20b. This marks the first time the company has launched open-source language models in over five years, since the release of GPT-2 in 2019. This move signifies a major strategic adjustment for OpenAI in response to fierce competition from Chinese AI labs.

Technical Specifications and Performance

gpt-oss-120b: Flagship Open-Source Model

  • Total Parameters: 117 billion parameters
  • Running Requirements: Can run on a single Nvidia GPU
  • Active Parameters: Only 5.1 billion parameters activated per token (utilizing Mixture-of-Experts architecture, MoE)
  • Performance: Scored 2622 in the Codeforces programming contest, surpassing DeepSeek R1

gpt-oss-20b: Lightweight Version

  • Parameter Count: 20 billion parameters
  • Hardware Requirements: Can run on a consumer-grade laptop with 16GB of RAM
  • Codeforces Score: 2516
  • Applicable Scenarios: Edge device deployment and rapid iterative development

Breakthroughs in Reasoning Capabilities

Both models possess Chain-of-Thought reasoning capabilities similar to OpenAI's 'o' series models, enabling them to:

  • Perform complex multi-step logical reasoning
  • Autonomously invoke tools such as web search and Python code execution
  • Excel in high-difficulty tasks like math competitions and programming challenges
  • Support low, medium, and high reasoning intensity modes

In authoritative tests, gpt-oss-120b matched or even surpassed OpenAI's o4-mini model on multiple benchmarks, while gpt-oss-20b, despite its smaller size, still performed excellently in competitive mathematics and health-related queries.

Market Competition Landscape

This release is not coincidental but a strategic response by OpenAI to intense market competition:

Rise of Chinese AI Labs

  • Chinese companies such as DeepSeek, Alibaba Qwen, and Moonshot AI have demonstrated outstanding performance in the open-source model domain
  • DeepSeek R1 achieved reasoning capabilities comparable to OpenAI's 'o' series at a lower cost
  • Meta's Llama series has seen its competitive position decline over the past year

Policy Environment Impact The Trump administration urged U.S. AI developers in July 2025 to open-source more technology to promote the global adoption of AI aligned with American values.

CEO Statement and Strategic Intent

OpenAI CEO Sam Altman stated in a declaration: "Back in 2015 when we founded OpenAI, our mission was to ensure AGI benefits all of humanity. To that end, we are excited to see the world building on an open AI technology stack created in the U.S. and based on democratic values."

Altman had previously admitted in January this year that OpenAI was "on the wrong side of history" regarding open-source technology. This release is seen as a significant correction to that stance.

Technical Implementation and Limitations

Training Methodology

  • Employs a training process similar to proprietary models
  • Utilizes high-compute reinforcement learning for post-training optimization
  • Integrates supervised fine-tuning and reinforcement learning techniques

Functional Limitations

  • Supports text input and output only, lacking image and audio processing capabilities
  • Hallucination Issue: In the PersonQA benchmark test, gpt-oss-120b and gpt-oss-20b exhibited hallucination rates of 49% and 53% respectively, significantly higher than the o1 model's 16%

Open-Source Degree

  • Uses the Apache 2.0 license, permitting commercial applications
  • Does not disclose training data to mitigate copyright legal risks
  • Only known components of the model architecture are open-sourced, protecting core proprietary technology

Safety Assessment

OpenAI conducted a comprehensive safety assessment of the models:

  • Potential marginal enhancement of biological capabilities, but not reaching the "high-risk" threshold
  • No significant increase in risk found concerning cyberattacks or chemical weapon manufacturing
  • Third-party evaluation agencies participated in safety testing

Deployment Convenience and Commercial Impact

gpt-oss-120b is suitable for enterprise applications requiring high-performance inference, with single-GPU deployment significantly lowering hardware barriers. gpt-oss-20b, on the other hand, offers small businesses and individual developers the possibility of running advanced AI inference models on consumer-grade hardware.

Both models are available for free download via the Hugging Face platform, and the Apache 2.0 license allows businesses to commercialize the models without requiring authorization from OpenAI.

Industry Impact and Outlook

This release is regarded by the industry as a significant turning point in the AI open-source race:

  1. Democratization of Technology: Lowers the barrier to entry for advanced inference AI
  2. Ecosystem Building: Lays the foundation for a U.S.-led open-source AI ecosystem
  3. Reshaping the Competitive Landscape: Directly challenges the market position of Chinese open-source models
  4. Acceleration of Innovation: Expected to drive rapid development of AI agent and tool-calling technologies

With the impending releases of DeepSeek R2 and Meta's new super-intelligent lab open-source models, the competition in the AI open-source domain will intensify in the second half of 2025. OpenAI's move is not only a correction of its past closed strategy but also an effort to re-establish its dominant position in setting global AI technology standards.