Workflow
Claude 小升级就赢了OpenAI 9年“开源神作”?高强度推理直接歇菜、幻觉率高达50%,写作还被Kimi 2吊锤?
AI前线·2025-08-06 04:25

Core Viewpoint - OpenAI has released its first open-source language model series, gpt-oss, which includes gpt-oss-120b and gpt-oss-20b, both of which are fully customizable and support structured output [2][3]. Model Specifications - gpt-oss-120b requires 80GB of memory to run, while gpt-oss-20b only needs 16GB [2]. - The models utilize a mixture of experts (MoE) architecture, activating 5.1 billion parameters per token for gpt-oss-120b and 3.6 billion for gpt-oss-20b, with total parameters of 117 billion and 21 billion respectively [9]. - Both models support a context length of up to 128k and are designed for efficient deployment on consumer-grade hardware [10]. Training and Performance - The training process for gpt-oss models combines reinforcement learning and techniques from OpenAI's advanced internal models, focusing on reasoning capabilities and efficiency [8]. - gpt-oss models have shown strong performance in reasoning tasks, with gpt-oss-120b performing comparably to OpenAI's proprietary models in core inference benchmarks [10]. Comparison with Competitors - Claude Opus 4.1 has demonstrated superior programming performance with a score of 74.5% in SWE-bench Verified programming evaluations, outperforming previous versions [5]. - Independent benchmark tests indicate that gpt-oss-120b is less intelligent than DeepSeek R1 and Qwen3 235B, although it has advantages in efficiency due to its smaller parameter size [13]. User Feedback and Limitations - Users have reported mixed experiences with gpt-oss models, noting that gpt-oss-120b is particularly unstable for coding tasks, while gpt-oss-20b performs better [6][17]. - The models exhibit a higher hallucination rate, with gpt-oss-120b and gpt-oss-20b generating hallucinations at rates of 49% and 53% respectively, significantly higher than OpenAI's previous models [16]. Open Source and Accessibility - gpt-oss models are released under the flexible Apache 2.0 license, making them accessible for various applications, including agent workflows and tool usage [11][10]. - The models are available for free download on Hugging Face, promoting wider adoption and experimentation within the developer community [2][3].