Google open-source Gemma 3
Here are the contents translated into English:
Gemini 3 is Google’s latest AI model, which enhances its language coverage, security, and performance through various techniques. The model uses four components during training: pre-trained checkpoints, reinforcement learning based on human feedback, machine feedback reinforcement learning, and execution feedback reinforcement learning. These methods significantly improve the model’s ability in mathematics, programming, and command following.
Gemini 3 supports multiple deployment environments, including Vertex AI, Cloud Run, Google GenAI API, local environments, and other platforms, allowing for the best solution to be chosen based on the application and infrastructure.
The Gemini 3 model has also been optimized for various hardware platforms, such as Jetson Nano and Blackwell chips, and can run on CPU. It has also improved security and privacy through strict filtering, SFT, and RLHF techniques to reduce the possibility of generating harmful content during pre-training and fine-tuning models.
However, Gemini 3 still has some limitations, such as performance that may not be comparable to professional customized models when handling extremely complex multimodal tasks. Despite this, the release of Gemini 3 marks an important step for Google in AI open source, expanding the possibilities for future AI applications on consumer-level hardware devices.
Translation
Gemini 3是谷歌最新推出的AI模型，通过多种技术提高其语言覆盖范围、安全性和性能。该模型在训练阶段使用了四个组件：预训练检查点、基于人类反馈的强化学习、机器反馈强化学习和强化学习执行反馈，这些方法显著提升了模型在数学、编程和指令跟随方面的能力。Gemma 3支持多种部署环境，包括Vertex AI、Cloud Run、Google GenAI API、本地环境和其他平台，可以根据应用和基础设施选择最佳方案。
Gemini 3还针对不同硬件进行了优化，如Jetson Nano、Blackwell芯片等，并且可以在CPU上执行。它对安全和隐私也进行了多项改进，通过严格的安全过滤、SFT和RLHF技术来降低预训练和微调模型生成有害内容的可能性。在隐私方面，它的长文本记忆率显著低于先前的模型。
然而，Gemma 3仍然存在一些不足，如处理某些极端复杂的多模态任务时性能可能不如专业定制模型等。尽管如此，Gemini 3的推出标志着谷歌在AI开源方面又迈出的重要一步，其多模态能力、长上下文处理以及良好适配性为未来AI在消费级硬件设备上的应用拓展了更为广阔的空间。
Reference:
https://blog.google/technology/developers/gemma-3/