OpenAI 推出两款开源模型，GPT-5蓄势待发！

Core Insights - OpenAI has launched two open-source models, GPT-oss-120b and GPT-oss-20b, marking its first release of open-source language models since GPT-2 in 2020 [2][5] - The models are available for free download on the Hugging Face platform and are reported to perform at the forefront of various benchmark tests [2][3] Model Performance and Specifications - Both models utilize advanced pre-training and post-training techniques, focusing on inference, efficiency, and practical deployment across environments [3] - GPT-oss-120b has a total parameter count of 117 billion, activating 5.1 billion parameters per token, while GPT-oss-20b has 21 billion total parameters, activating 3.6 billion parameters [3] - The models support a context length of up to 128k and are designed to run on high-end consumer GPUs and Apple chip-equipped devices [3] Competitive Landscape and Strategic Shift - OpenAI's shift to open-source models is a response to increasing market competition, particularly following the rise of open-source AI technologies [5] - The launch aims to attract more developers and enterprises into OpenAI's ecosystem, enhancing its competitive position in the AI sector [5] - Collaborations with chip manufacturers like NVIDIA and AMD are intended to ensure optimal performance of the models across various hardware [5] Safety and Market Expectations - Despite strong performance, the new models are more prone to "hallucination" phenomena compared to previous models, with hallucination rates of 49% for GPT-oss-120b and 53% for GPT-oss-20b [6] - OpenAI has implemented safety measures during pre-training to filter harmful data and assess potential risks [6] - There is growing market anticipation for the next major product from OpenAI, expected to be GPT-5, which aims to simplify and unify capabilities across its model series [6][7]