NKY(SZ:300109) 3 6 Ke·2025-08-06 00:55Core Insights - OpenAI has released two new open-source language models, gpt-oss-120b and gpt-oss-20b, marking a significant return to open-sourcing since GPT-2 [1][2] - The models are designed for reasoning tasks and can be run on consumer-grade hardware, with gpt-oss-120b performing comparably to OpenAI's proprietary models [5][14] Model Features - Both models are licensed under the Apache 2.0 license, allowing for unrestricted use and customization [7] - They offer adjustable reasoning capabilities to meet specific use cases and latency requirements [7] - The models provide a complete chain of thought (CoT) for easier debugging and increased trust in outputs, although this feature is not intended for end-user display [8] - They support fine-tuning for specific applications and include agentic functionalities for function calls, web browsing, and Python code execution [8] Technical Specifications - gpt-oss-120b has 117 billion total parameters with 5.1 billion active parameters, while gpt-oss-20b has 21 billion total parameters with 3.6 billion active parameters [17] - Both models utilize a transformer architecture and expert mixture (MoE) to optimize performance and reduce active parameters during input processing [16] - They support a maximum context length of 128k tokens and employ advanced techniques for training and post-training [16][18] Performance Evaluation - In various benchmark tests, gpt-oss-120b outperformed OpenAI's o3-mini and matched the performance of o4-mini in programming competitions and tool usage [19] - gpt-oss-20b also showed competitive performance, particularly in medical queries and competition math, despite its smaller size [19] - The models have been tested against academic benchmarks, demonstrating strong capabilities in programming, reasoning, and tool usage [19] Community Response - The release has generated significant interest in the AI community, leading to high download activity and server strain on platforms like Hugging Face [10]