六年来首次！OpenAI新模型开放权重，Altman称为“全球最佳开放模型”

Core Viewpoint - OpenAI has made a significant move by releasing its first open-weight language models, gpt-oss-120b and gpt-oss-20b, in six years, responding to competition in the open-weight model space [1][2][16]. Model Details - The newly released models, gpt-oss-120b and gpt-oss-20b, are designed for low-cost options, allowing developers and researchers to run and customize them easily [2][5]. - gpt-oss-120b has a total parameter count of 117 billion, with 5.1 billion active parameters per token, while gpt-oss-20b has 21 billion total parameters and 3.6 billion active parameters per token [6][7]. - Both models support a context length of up to 128k tokens and utilize a mixture of experts (MoE) architecture to enhance efficiency [5][6]. Performance Evaluation - gpt-oss-120b's performance in core inference benchmarks is comparable to OpenAI's o4-mini, while gpt-oss-20b matches or exceeds the performance of o3-mini in similar evaluations [8][9]. - The models are optimized for local execution, with gpt-oss-20b requiring at least 16GB of memory and gpt-oss-120b needing around 80GB [2][8]. Strategic Shift - The release marks a strategic shift for OpenAI, which had previously delayed the launch of open-weight models due to extensive safety testing [3][16]. - OpenAI's decision to release these models is largely driven by competitive pressure, particularly from companies like DeepSeek [16]. Security Measures - OpenAI has implemented rigorous security measures to prevent malicious use of the models, including filtering harmful data during pre-training and conducting extensive internal and external testing [14][15]. - The company has collaborated with independent expert groups to assess potential security risks associated with the models [15]. Ecosystem and Market Impact - The models are available under the Apache 2.0 license and can be downloaded from platforms like Hugging Face and GitHub, with cloud service providers like Amazon and Microsoft also offering them [15][16]. - OpenAI is hosting a red team challenge with a $500,000 prize pool to encourage the identification of new security issues [15].