来源：互联网2026-01-15 00:00:00 热度：

田丰：国产AI反封锁，智谱AI开源基于腾芯片训练的多模态模型

AI中国网 https://www.cnaiplus.com

导读：观点摘要快思慢想研究院院长、特邀评论员田丰认为GLM-Image的发布，是一次典型的技术驱动型突破，其核心价值在于揭示了多模态AI发展的三条深层技术路径：1）架构范式的创新性重构： GLM-Image采用的“自回归+扩散”混合架构，绝非简单叠加。其本质是将视觉生成的认知过程解耦为“理解”与“渲染”两大任务，分别交由自回归模块与扩散模块处理。这不仅实现了模型“左脑懂逻辑、 ......

观点摘要

快思慢想研究院院长、特邀评论员田丰认为GLM-Image的发布，是一次典型的技术驱动型突破，其核心价值在于揭示了多模态AI发展的三条深层技术路径：

1）架构范式的创新性重构： GLM-Image采用的“自回归+扩散”混合架构，绝非简单叠加。其本质是将视觉生成的认知过程解耦为“理解”与“渲染”两大任务，分别交由自回归模块与扩散模块处理。这不仅实现了模型“左脑懂逻辑、右脑画细节”的协同，更验证了一条不依赖单纯堆叠算力，而是通过算法架构创新实现性能跃升的可行路径。这标志着AI模型正在从“拼接式”融合，迈向“原生一体化”的认知生成。

2）国产算力自主的工程性验证：本次模型是首个在国产腾算力与MindSpore框架上完成全程训练的SOTA模型，其意义超越了性能本身。它从数据并行、算子优化到通信调度，系统性验证了国产AI全栈（芯片-框架-模型）支撑前沿大模型训练的工程可行性。成本优势（单图0.1元）证明了其在实现“自主可控”的同时，已具备商业化落地所需的“性价比”竞争力，为中国AI构建不依赖于单一技术体系的长远发展提供了实证案例。

3）技术范式的代际前瞻：该模型是智谱对“认知型生成”范式的探索，旨在推动AI从“感知像素”的模仿工具，迈向“理解意图、具备知识”的认知智能体。这意味着AI生成任务的核心将从“怎么画得像”转向“如何理解对”，这将是AI向更高阶通用智能演进的重要转折点。

中文翻译版

作者：《环球时报》张蔚蓝

田丰：国产AI反封锁，智谱AI开源基于腾芯片训练的多模态模型

智谱AI本周三（2026年1月14日）宣布，与华为合作开源新一代图像生成模型GLM-Image。该模型代表了多模态AI领域的先进水平里程碑。

智谱表示，GLM-Image模型是国内首个完全使用国产芯片完成训练的多模态AI模型。

有专家指出，GLM-Image的开源彰显了中国AI开发者和硬件供应商在构建覆盖芯片、框架与大模型的自主AI技术栈方面的持续努力。

该公司周三告知《环球时报》，GLM-Image实现了端到端训练从数据处理到模型训练均使用华为腾Atlas 800T A2硬件，并运行在MindSpore AI框架上，成为据报道首个在国产芯片上训练达到先进性能的开源多模态模型。

公司表示，这一进展标志着中国AI产业在技术自主可控的道路上迈出坚实一步，证明了在国产全栈计算平台上训练高性能多模态生成模型的可行性。

据智谱AI介绍，GLM-Image采用混合式“自回归+扩散解码器”架构，有别于常用的潜在扩散模型（LDM）方法。公司称，这种新范式能实现语言与图像生成的更紧密融合，并在知识密集型生成场景中取得更佳效果。

智谱AI一位研究员周三对《环球时报》表示，通过与华为紧密合作，团队在腾Atlas 800T A2设备上完成了从数据准备到大模型训练、推理适配的全流程，经联合调试优化后，训练性能已接近目标硬件的实用极限。

“GLM-Image的目标从一开就是全栈创新，”智谱AI研究员郑文迪说，“我们验证了新的自回归+扩散解码器架构，并在腾Atlas 800T A2设备上实现了完整的训练与推理适配。华为提供了及时的调试和性能优化支持，帮助我们解决了诸多瓶颈。”

GLM-Image也具备优越的商业成本表现。它将图像生成与大语言模型能力融合，可实现统一的多模态输出。在API使用模式下，生成单张图片成本仅0.1元人民币（0.014美元）。公司表示，预计不久将发布优化后的更快版本模型。

“此次合作标志着国产AI软硬件生态在核心技术上的重大进展，为国产芯片能够处理复杂AI任务提供了重要验证，”快思慢想研究院院长、原商汤智能产业研究院创始院长田丰表示。他补充说，这也表明尽管面临外部技术封锁，中国自主技术创新的动力依然强劲。

田丰周三对《环球时报》表示，短期来看，这一进展预计将提振国内AI产业链信心，直接惠及腾芯片、腾框架相关企业及智谱生态伙伴。

长期而言，技术自主进程的加速或将重塑AI计算格局，降低对外国硬件的依赖，并激发全国范围内的创新。但田丰也指出，商业化进展、日益激烈的海外竞争以及算力供应链的稳定性，仍需持续努力。

被OpenAI公开视为竞争对手的智谱，已成为首批上市的中国AI公司之一，标志着中国国产AI模型正从技术探索走向大规模商业应用。自此，其股价已上涨逾80%，投资者对中国AI产业的热情持续高涨。

专家指出，中国企业的自主创新正开始克服过去几年美国单边技术封锁造成的困难。此外，中国企业将继续增强自身韧性，并加强研发能力。

据新华社2025年8月报道，在智谱此举之前，华为已宣布将其腾芯片软件生态全面开源，旨在支持用户深入挖掘其潜力并独立进行定制化开发。

新华社报道，华为表示，经过多年发展，其软件系统CANN（计算架构神经网络）已在计算优化、通信效率和内存管理方面实现关键突破，现已能够为AI模型训练和部署的全流程提供算力支持。

英文原文

Chinese AI startup Zhipu AI announced on Wednesday that it has partnered with Huawei to open-source GLM-Image, a new-generation image generation model that represents a state-of-the-art (SOTA) milestone in multimodal AI.Zhipu said the GLM-Image model is the country's first multimodal AI model to be fully trained using domestically produced chips.The open-sourcing of GLM-Image highlights ongoing efforts by Chinese AI developers and hardware providers to build a self-sustaining AI technology stack, spanning chips, frameworks, and large-scale models, a Chinese expert said.GLMImage was trained end to end from data processing to model training using Huawei's Ascend Atlas 800T A2 hardware and running on the MindSpore AI framework, making it the first opensource multimodal model reported to reach SOTA performance after being trained on domestically developed Chinese chips, the company told the Global Times on Wednesday.The progress marks a solid step forward for China's AI industry on the path toward independent and controllable technology. This demonstrates the feasibility of training high-performance multimodal generative models on a domestically developed full-stack computing platform, the company said.According to Zhipu AI, GLMImage uses a hybrid "autoregressive + diffusion decoder" architecture that departs from the commonly used latent diffusion model (LDM) approach. The company said the new paradigm enables tighter integration between language and image generation and delivers improved results in knowledgeintensive generation scenarios. A research fellow at Zhipu AI told the Global Times on Wednesday that through close collaboration with Huawei, the team completed the full pipeline from data preparation to largescale training and inference adaptation on Ascend Atlas 800T A2 devices, with training performance approaching the practical limits of the targeted hardware after joint debugging and optimization. "From the beginning, GLMImage's goal was fullstack innovation," said Zheng Wendi, a research fellow at Zhipu AI. "We validated a new autoregressive + diffusion decoder architecture and implemented a complete training and inference adaptation on Ascend Atlas 800T A2 devices. Huawei provided timely debugging and performance-optimization support, helping us address many bottlenecks.GLM-Image also offers a strong commercial cost profile. It integrates image generation with large language model capabilities, enabling unified multimodal outputs. Under an API-based usage model, generating a single image costs just 0.1 yuan ($0.014). An optimized, faster version of the model is expected to be released soon, the company said."This collaboration marks a significant advance in core technologies for the domestic AI software and hardware ecosystem, offering important validation that domestically produced chips can handle complex AI tasks," said Tian Feng, president of the Fast Think Institute and former dean of SenseTime's Intelligence Industry Research Institute. He added that it also shows China's drive for independent technological innovation has continued unabated despite external technology blockades. In the short term, the progress is expected to boost confidence across the local AI industry chain, directly benefiting companies tied to Ascend chips and the Ascend framework, as well as partners in the Zhipu ecosystem, Tian told the Global Times on Wednesday.Over the long term, accelerated technological self-reliance could reshape the AI computing market, reduce dependence on foreign hardware, and spur nationwide innovation. However, commercialization progress, intensifying overseas competition, and the stability of the computingpower supply chain will still require sustained effort, Tian said.Zhipu, which OpenAI publicly identifies as a rival, has become one of the first Chinese AI firm togo public, as China's homegrown AI models move from technological exploration to large-scale commercial application. Since then, its shares have jumped more than 80 percent as investors pile in on enthusiasm about China's AI industry. Indigenous innovation by Chinese companies is beginning to overcome the difficulties caused by the US unilateral technology blockade over the past few years. Moreover, Chinese companies are poised to continue enhancing their resilience and strengthening their research and development capabilities, said the expert.Zhipu's move came after Huawei previously announced the full open-source release of its Ascend chip software ecosystem, aiming to support users to explore its deep potential and undertake customized development independently, the Xinhua News Agency reported in August, 2025.Huawei said that after years of development, the company's software system Compute Architecture for Neural Networks has achieved key breakthroughs in computing optimization, communication efficiency and memory management. It is now capable of providing computing power support throughout the entire AI-model training and deployment process, Xinhua reported.

网址：https://enapp.globaltimes.cn/article/1353199

该图片可能由AI生成田丰：国产AI反封锁，智谱AI开源基于腾芯片训练的多模态模型