gpt-oss - OpenAI's open-source inference model series

What is GPT-OSS?

gpt-oss is an open-source inference model series launched by OpenAI, including two versions: gpt-oss-120b and gpt-oss-20b. gpt-oss-120b has 117 billion parameters, with approximately 5.1 billion activation parameters, and can run on a single 80GB GPU; gpt-oss-20b has 21 billion parameters, with approximately 3.6 billion activation parameters, and can run on consumer-grade devices (such as laptops or mobile phones) with 16GB of RAM. Both are based on the MoE architecture, support a 128k context length, and their performance is close to that of the closed-source o4-mini and o3-mini, demonstrating excellent performance in tool calls, few-shot function calls, chained reasoning, and health question answering. The models are open-source under the Apache 2.0 license, supporting free commercial use and providing developers with powerful local inference capabilities.

gpt-oss-safeguard isOpenAI's latest open security inference model, supporting custom security policies, includes two versions: gpt-oss-safeguard-120b and gpt-oss-safeguard-20b. These models are specifically designed for content reasoning, classification, and labeling, and are fine-tuned versions of the gpt-oss model. Both models support complete "thought chain" outputs, demonstrating the reasoning process. They can quickly adapt to new rules without retraining, offering high flexibility and suitability for scenarios involving emerging risks, complex domains, and insufficient sample sizes.

Main functions of gpt-oss

• Tool callIt supports calling external tools (such as web search and Python code interpreter) to solve complex problems, enhancing problem-solving capabilities.
• Chain thinkingIt breaks down complex problems step by step and solves them step by step, making it suitable for multi-step reasoning tasks.
• Low resource demandgpt-oss-20b can run on devices with 16GB of memory, making it suitable for consumer devices; gpt-oss-120b can run on GPUs with 80GB of memory, making it suitable for high-performance needs.
• Rapid reasoningThe inference speed can reach 40-50 tokens/s, making it suitable for scenarios that require fast response.
• Open source weightIt provides complete model weights and code, supports local fine-tuning and customization, and meets specific task requirements.
• Reasoning intensity adjustmentIt supports three inference intensities: low, medium, and high, which can be adjusted according to needs to balance latency and performance.
• gpt-oss-safeguard:
  • Content Reasoning and ClassificationBased on the strategies provided by the developers, infer, classify, and label user messages or AI-generated content.
  • Dynamic strategy applicationIt directly interprets and applies developer-defined security policies during the inference phase, eliminating the need to hardcode rules during the training phase and enabling rapid adaptation to new rules.
  • Complete "mind chain" outputIt demonstrates each step of the reasoning process that leads the model to a conclusion, improving transparency and usability, and making it easier for developers to trace and understand the model's judgment logic.

The technical principles of gpt-oss

• Model ArchitectureBased on the Transformer architecture, it uses Hybrid Expert (MoE) technology to reduce the number of active parameters required to process the input, thereby improving inference efficiency. It further optimizes memory and computation efficiency by employing alternating dense and locally striped sparse attention patterns, similar to GPT-3. A grouped multi-query attention mechanism with a group size of 8 further enhances inference efficiency. RoPE is used for positional encoding, supporting context lengths up to 128k.
• Pre-training and post-trainingPre-training utilizes high-quality plain text datasets, focusing on STEM, programming, and general knowledge domains. Post-training employs a similar workflow to o4-mini, including supervised fine-tuning and computationally intensive reinforcement learning phases. The training objective is to ensure the model conforms to OpenAI model specifications, enabling chained inference and tool invocation capabilities.
• Quantification and OptimizationQuantization is performed using the MXFP4 format, allowing the model to adapt to low-precision environments during training, ensuring high performance while reducing model size. The gpt-oss-20b quantized model is approximately 12.8GB in size and can run on devices with 16GB of RAM. The gpt-oss-120b quantized model can run on devices with 80GB of RAM. Collaboration with hardware manufacturers such as NVIDIA and AMD ensures performance optimization across various systems.
• Security MechanismDuring the pre-training phase, harmful data related to chemistry, biology, radioactivity, and nuclear (CBRN) is filtered. Through careful alignment and instruction prioritization evaluation, the trained model rejects unsafe cues and resists injection attacks. Adversarial fine-tuning is performed in specific domains (such as biology and cybersecurity) to evaluate and optimize the model's security.

gpt-oss performance

• Benchmark performance:
  • Competition ProgrammingIn the Codeforces programming competition, gpt-oss-120b scored 2622 points and gpt-oss-20b scored 2516 points, outperforming some open-source models but slightly underperforming the closed-source o3 and o4-mini.
  • General Problem SolvingIn MMLU and HLE tests, gpt-oss-120b outperformed OpenAI's o3-mini and approached the level of o4-mini.
  • Tool callIn the TauBench agent evaluation kit, gpt-oss-120b and gpt-oss-20b both outperformed OpenAI's o3-mini and reached or exceeded the level of o4-mini.
  • Health Q&AIn HealthBench tests, gpt-oss-120b outperformed o4-mini, while gpt-oss-20b reached a level comparable to o3-mini.

gpt-oss project address

gpt-oss:

• Project official website: https://openai.com/zh-Hans-CN/index/introducing-gpt-oss/
• GitHub repositoryhttps://github.com/openai/gpt-oss
• HuggingFace model library: https://huggingface.co/collections/openai/gpt-oss-68911959590a1634ba11c7a4
• Experience the demo onlinehttps://gpt-oss.com/

gpt-oss-safeguard:

• Project official websitehttps://openai.com/index/introducing-gpt-oss-safeguard/
• HuggingFace model libraryhttps://huggingface.co/collections/openai/gpt-oss-safeguard
• Technical Papers: https://openai.com/index/gpt-oss-safeguard-technical-report/

Application scenarios of gpt-oss

• Local Inference and Privacy ProtectionIn privacy-sensitive fields (such as healthcare and finance), gpt-oss can run on local devices to ensure data security while providing powerful inference capabilities.
• Programming assistance and code generationDevelopers can use gpt-oss to call tools to generate and verify code snippets, significantly improving programming efficiency and reducing debugging time.
• Intelligent Customer Service and Customer SupportEnterprises can deploy gpt-oss as an intelligent customer service system to quickly respond to customer inquiries, provide accurate answers, and reduce labor costs.
• Education and Learning GuidanceStudents use gpt-oss to assist their learning, answer questions, and receive writing suggestions, thereby improving their learning efficiency and comprehension.
• Creative content generationWriters, screenwriters, game developers, and others use gpt-oss to generate creative content, inspire ideas, and improve creative efficiency.