أوبن إيه آي تطلق نماذج لغوية مفتوحة المصدر رائدة: gpt-oss-120b و gpt-oss-20b

OpenAI has announced the release of two new open-source language models, gpt-oss-120b and gpt-oss-20b, marking a significant stride in the accessibility and advancement of AI technology. These models aim to deliver strong performance at a lower cost, enabling wider adoption and customization across various applications. Available under the Apache 2.0 license, they outperform similarly sized open models on reasoning tasks, demonstrate robust tool use capabilities, and are optimized for efficient deployment on consumer hardware.

Key Features and Capabilities

• High Performance: gpt-oss-120b closely matches OpenAI’s o4-mini on reasoning benchmarks, while the smaller gpt-oss-20b mirrors the performance of o3-mini.
• Resource Efficiency: gpt-oss-120b can operate on a single 80 GB GPU, and gpt-oss-20b is designed for edge devices with just 16 GB of memory.
• Tool Use and Reasoning: Both models excel in tool use, few-shot function calling, and chain-of-thought (CoT) reasoning, as demonstrated by their performance on benchmarks like Tau-Bench and HealthBench.
• Customization and Control: The models support full chain-of-thought (CoT), structured outputs, and are entirely customizable, providing developers with granular control over their behavior.
• Compatibility: These models are designed to work seamlessly with the Responses API and within agentic workflows.

Architecture and Training

• Mixture-of-Experts (MoE): Both models utilize a Transformer architecture that leverages MoE to reduce the number of active parameters per token (5.1B for gpt-oss-120b and 3.6B for gpt-oss-20b).
• Attention Patterns: The models employ alternating dense and locally banded sparse attention patterns, drawing inspiration from GPT-3.
• Grouped Multi-Query Attention: For enhanced inference and memory efficiency, grouped multi-query attention with a group size of 8 is used.
• Rotary Positional Embedding (RoPE): RoPE is used for positional encoding, supporting context lengths up to 128k.
• Training Data: The models were trained on a primarily English, text-only dataset focused on STEM, coding, and general knowledge, tokenized using a superset of the tokenizer used for o4-mini and GPT-4o.

Post-Training and Alignment

• Supervised Fine-Tuning & Reinforcement Learning: A similar post-training process as used for o4-mini was employed, including supervised fine-tuning and reinforcement learning to align the models with the OpenAI Model Spec.
• Reasoning Efforts: The models support three reasoning efforts (low, medium, high) to balance latency and performance.

Safety Measures

• Comprehensive Safety Training: The models underwent extensive safety training and evaluations.
• Adversarial Fine-Tuning Evaluation: An adversarially fine-tuned version of gpt-oss-120b was tested under OpenAI’s Preparedness Framework.
• Internal Safety Benchmarks: gpt-oss models perform comparably to OpenAI’s frontier models on internal safety benchmarks.
• Red Teaming Challenge: A $500,000 Red Teaming Challenge is being hosted to identify novel safety issues, with a report and evaluation dataset to be open-sourced.

Availability and Integration

• Hugging Face: The model weights are available for free download on Hugging Face in MXFP4.
• Hardware Optimization: Optimized for NVIDIA, AMD, Cerebras, and Groq hardware.
• Deployment Platforms: Partners include Azure, Hugging Face, vLLM, Ollama, llama.cpp, LM Studio, AWS, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare, and OpenRouter.
• Microsoft Integration: GPU-optimized versions of gpt-oss-20b are available on Windows devices via ONNX Runtime, Foundry Local, and the AI Toolkit for VS Code.

OpenAI hopes that the release of gpt-oss-120b and gpt-oss-20b will empower developers, researchers, and organizations worldwide to build and customize AI solutions on their own infrastructure, fostering innovation and democratizing access to advanced AI technology. These open models are poised to accelerate leading-edge research, foster innovation, and enable safer, more transparent AI development across a wide range of use cases.

Source: N/A