[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1HrgeX5y5DD1TgwEu13YM6AdI8nUh9lxZLl-7G2IggU":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},2061,"谷歌推出DiffusionGemma文本扩散AI模型，本地推理效率实现大幅提升","\u003Cp>\u003Cspan style=\"font-size: 18px;\">6 月 11 日，谷歌正式推出全新文本扩散 AI 模型 DiffusionGemma，该模型凭借创新架构将本地推理速度提升至传统自回归模型的 4 倍，为端侧及本地 AI 应用发展开辟新方向。目前 GPT、Gemini 等主流大模型均采用自回归架构，需按顺序逐一生成文本单元，虽适配云端作业，但在本地运行时易受内存带宽制约，造成算力损耗。DiffusionGemma 另辟蹊径，运用文本扩散技术，以噪声逐步去噪、并行处理文本单元的方式生成内容，高度适配本地低带宽环境。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">该模型综合能力与同系列 Gemma 4 模型持平，兼顾生成质量与运行效率，支持迭代纠错，输出内容稳定性更强。其采样速率可达每秒 1479 个文本单元，单次生成仅需 0.84 秒。多项测试展现出差异化实力，代码生成领域表现亮眼，在三项权威测评中分别取得 30.9%、45.4%、89.6% 的成绩，整体水平比肩 Gemini 2.0 Flash-Lite；数学推理能力突出，在 AIME 2025 测试中得分 23.3%，优于同类对比模型。不过该模型仍存在短板，在科学推理、高难度综合推理场景中表现不及主流模型。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">硬件适配层面，DiffusionGemma 可充分发挥英伟达 GPU 的并行计算能力。实测数据显示，不同设备运行速度差异明显，单块 H100 GPU 每秒可生成 1000 个文本单元，DGX Spark 设备速率为每秒 150 个，DGX Station 设备可达每秒 2000 个文本单元。现阶段，DiffusionGemma 已依据 Apache 2.0 开源协议对外开放，开发者可前往 Hugging Face 平台下载模型权重，开展二次开发与落地实践。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>【新闻来源】环球网  https:\u002F\u002Ftech.huanqiu.com\u002Farticle\u002F4RvtGkCfahP\u003C\u002Fp>\u003Cp>（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2026\u002F06\u002F48323942b2c745faa7c477e998e1432b\u002F芯位视野.png","https:\u002F\u002Fimage.51xinwei.com\u002F2026\u002F06\u002Fthumbs\u002F48323942b2c745faa7c477e998e1432b\u002F芯位视野.png",0,1,45,"2026-06-12 15:06",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A2aa5fd22-40c8-40bd-af97-f77eb0b9d283%3A0.wav?Expires=1781253968&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=H3xSZqbeTXrsJm6G4wp4hd7fxRs%3D",3768818,"2aa5fd22-40c8-40bd-af97-f77eb0b9d283","2026-06-12 15:02","Google Launches DiffusionGemma Text Diffusion AI Model, Achieving Significant Improvement in Local Inference Efficiency","\u003Cp>\u003Cspan style=\"font-size: 18px;\">On June 11th, Google officially launched the brand-new text diffusion AI model, DiffusionGemma. Leveraging an innovative architecture, this model has increased local inference speed to four times that of traditional autoregressive models, paving a new direction for the development of on-device and local AI applications. Currently, mainstream large models such as GPT and Gemini all employ autoregressive architectures, which require generating text units sequentially one by one. While suitable for cloud-based tasks, they are prone to memory bandwidth constraints during local operation, leading to computational waste. DiffusionGemma takes a different path, utilizing text diffusion technology to generate content by gradually denoising from noise and processing text units in parallel, making it highly suitable for low-bandwidth local environments.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The model's overall capabilities are on par with its sibling model, Gemma 4, balancing generation quality and operational efficiency, and supports iterative error correction, resulting in stronger output stability. Its sampling rate can reach 1479 text units per second, with a single generation taking only 0.84 seconds. Multiple tests demonstrate differentiated strengths; its performance in code generation is particularly notable, achieving scores of 30.9%, 45.4%, and 89.6% in three authoritative evaluations respectively, with an overall level comparable to Gemini 2.0 Flash-Lite. Its mathematical reasoning ability is also outstanding, scoring 23.3% in the AIME 2025 test, outperforming similar comparative models. However, the model still has shortcomings, performing less well than mainstream models in scientific reasoning and high-difficulty comprehensive reasoning scenarios.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">In terms of hardware compatibility, DiffusionGemma can fully utilize the parallel computing capabilities of NVIDIA GPUs. Actual measurement data shows significant speed differences across different devices. A single H100 GPU can generate 1000 text units per second, DGX Spark devices achieve a rate of 150 per second, and DGX Station devices can reach 2000 text units per second. At this stage, DiffusionGemma has been made publicly available under the Apache 2.0 open-source license. Developers can download the model weights from the Hugging Face platform for secondary development and practical implementation.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>[News Source] Global Network https:\u002F\u002Ftech.huanqiu.com\u002Farticle\u002F4RvtGkCfahP\u003C\u002Fp>\u003Cp>(This website reposts this article to provide readers with more information. The content involved does not constitute investment or consumption advice. If there are questions regarding the facts in the article, please verify with the relevant parties. The views in the article are not the views of this website and are for readers' reference only.)\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A835fb24f-f73e-4d1b-89ef-f0507abced27%3A0.wav?Expires=1781253968&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=%2BI0FYfztzippE3%2BWveVfL3zvyZQ%3D","835fb24f-f73e-4d1b-89ef-f0507abced27",5391516]