2月26日，摩尔线程官微宣布已在AI训推一体全功能GPU MTT S5000上，完成对阿里三款全新模型的全 方位适配。据介绍，继开源Qwen3.5-397B-A17B之后，阿里宣布开源千问3.5最新三款中等规模模型 Qwen3.5-35B-A3B、Qwen3.5-122B-A10B、Qwen3.5-27B。 ...

2月24日，马年开工首日 "物理AI第一股"五一视界 宣布达成深度技术合作！ 海淀"国产GPU第一股"摩尔线程和 双方将基于 摩尔线程的全功能GPU MTT S5000与 五一视界的下一代智驾仿真平台SimOne 4.0 共同打造全栈国产化的物理AI仿真体系 为智能驾驶行业 提供自主可控的核心算力与仿真解决方案 据介绍，随着端到端智驾路线向世界模型与VLA（视觉-语言-动作）模型收敛，如何高效处理"长尾场 景"、实现高置信度的闭环仿真与合成数据生成，已成为行业公认的技术瓶颈。为破解这一关键难题， 摩尔线程与五一视界强强联合，打通国产全功能GPU算力与端到端智能驾驶仿真体系，构建了从底层芯 片到上层应用的国产软硬件协同闭环。 国产算力底座，打通仿真关键链路 本次合作的核心，是基于摩尔线程旗舰级全功能GPU MTT S5000，对五一视界SimOne 4.0平台进行系统 性适配与深度优化。双方成功贯通了大模型感知挖掘、4DGS（4D高斯泼溅）模型训练、4DGS仿真推 理及合成数据生成等关键链路，开启了全栈国产化高置信度仿真的新篇章。 图示：基于SimOne 4.0实现7个周视摄像头4DGS场景重建效果 MTT S ...

2026年新版《道路机动车辆生产企业准入审查要求》中，仿真验证能力被作为强制性关键要素写入准入 审查要求，国产仿真平台迎来巨大的市场利好，进一步推动了智驾产业链加速向"自主可控"迈进。 在此背景下，五一视界宣布旗下51Sim下一代智驾仿真平台 SimOne 4.0 已于近日成功完成在摩尔线程旗 舰级AI训推一体全功能GPU MTT S5000上的系统性适配与深度优化。从大模型感知挖掘、4DGS模型训 练到4DGS仿真推理和合成数据生成，双方强强联手，实现了国产算力与端到端智能驾驶仿真体系的完 整打通，正式开启了物理AI智驾领域高置信度闭环仿真与合成数据的全栈国产化新篇章,为应对新规下 的严苛验证需求提供了坚实的国产底座。 (一)算力底座的"跨代式"追赶 随着端到端智驾收敛到VLA和世界模型两条路线，基于海量Log数据的高置信度闭环仿真和合成数据已 成为提升算法长尾场景处理能力的关键命题。 仿真效率的上限，往往取决于基础一硬件的规格。在很长一段时间里，高端GPU市场几乎被海外巨头垄 断，但五一视界与摩尔线程的最新合作成果显示，这一格局正在发生微妙的变化。 在实测数据中，MTT S5000展现出了足以抗衡甚至部分 ...

据介绍，双方的强强联手，首次将国产算力与端到端智能驾驶仿真体系完整打通，标志着从芯片、仿真 引擎到上层应用的国产软硬件生态闭环正加速形成，共同构建物理AI进化的基础设施。未来，双方将 持续深化技术融合，打磨国产算力底座上的算法表现，共建繁荣的智驾创新生态，为全球自动驾驶产业 输出具备核心竞争力的"中国方案"。#摩尔线程# 中经记者 李玉洋 上海报道 【#摩尔线程与五一视界共建物理AI仿真国产化生态#】2月24日，"国内GPU第一股"摩尔线程 （688795.SH）宣布与五一视界共建全栈国产化的物理AI仿真体系，摩尔线程以旗舰级AI训推一体全功 能GPU MTT S5000的强劲算力，深度赋能五一视界下一代智驾仿真平台SimOne 4.0，双方已高效完成系 统适配与深度优化。 "当前随着端到端智驾路线收敛至VLA与世界模型，提升算法长尾场景处理能力成为行业尚未破解的关 键瓶颈，基于海量Log数据的高置信度闭环仿真与合成数据生成更是行业公认的技术难点。"摩尔线程 方面对《中国经营报》记者表示，为突破这一挑战，摩尔线程联合五一视界，通过打通从大模型感知挖 掘、4DGS模型训练到4DGS仿真推理和合成数据生成的关键链 ...

据介绍，双方的强强联手，首次将国产算力与端到端智能驾驶仿真体系完整打通，标志着从芯片、仿真 引擎到上层应用的国产软硬件生态闭环正加速形成，共同构建物理AI进化的基础设施。未来，双方将 持续深化技术融合，打磨国产算力底座上的算法表现，共建繁荣的智驾创新生态，为全球自动驾驶产业 输出具备核心竞争力的"中国方案"。 （文章来源：中国经营报） 2月24日，"国内GPU第一股"摩尔线程（688795.SH）宣布与五一视界共建全栈国产化的物理AI仿真体 系，摩尔线程以旗舰级AI训推一体全功能GPU MTT S5000的强劲算力，深度赋能五一视界下一代智驾 仿真平台SimOne 4.0，双方已高效完成系统适配与深度优化。 "当前随着端到端智驾路线收敛至VLA与世界模型，提升算法长尾场景处理能力成为行业尚未破解的关 键瓶颈，基于海量Log数据的高置信度闭环仿真与合成数据生成更是行业公认的技术难点。"摩尔线程 方面对《中国经营报》记者表示，为突破这一挑战，摩尔线程联合五一视界，通过打通从大模型感知挖 掘、4DGS模型训练到4DGS仿真推理和合成数据生成的关键链路，基于MTT S5000，正式开启物理AI高 置信度闭环仿真与合成 ...

北京商报讯（记者 陶凤 王天逸）2月24日，摩尔线程、五一视界宣布，以摩尔线程旗舰级AI训推一体 全功能GPU MTT S5000的强劲算力，深度赋能五一视界下一代智驾仿真平台SimOne 4.0，双方已高效完 成系统性适配与深度优化。 据称，通过打通从大模型感知挖掘、4DGS模型训练到4DGS仿真推理和合成数据生成的关键链路，双方 基于MTT S5000，正式开启物理AI高置信度闭环仿真与合成数据的全栈国产化新篇章。 ...

Group 1 - The core achievement of Moore Threads is the full adaptation of Alibaba's latest large model Qwen3.5 on its flagship AI training and inference GPU MTT S5000, enabling efficient model deployment and optimization through MUSA C programming language and Triton-MUSA toolchain [2] - The adaptation process validated two core capabilities of the MUSA ecosystem: native MUSA C support allows developers to directly use MUSA C for kernel development, significantly lowering the migration barrier from CUDA; deep compatibility with Triton-MUSA enables developers to write high-performance operators using familiar Triton syntax, seamlessly running on Moore Threads' full-featured GPU [2] - Moore Threads achieved native optimization for the hybrid attention mechanism used in the Qwen3.5 multimodal model, providing efficient support for long sequence processing based on the muDNN computing library and MATE open-source operator library, successfully realizing high-performance inference of the model on the MTT S5000 [2] Group 2 - In a week, Moore Threads rapidly followed up and adapted to top domestic large models such as GLM-5, MiniMax M2.5, and Qwen3.5, indicating a normalized agile response mechanism [3] - This agile response is attributed to the seamless compatibility of the MUSA architecture with mainstream AI ecosystems and the continuous optimization of the toolchain, marking the establishment of a full-link support capability from model adaptation to efficient deployment for domestic computing power [3] - A new ecological model combining domestic computing power and domestic large models is gradually taking shape [3]

Group 1 - Tencent's cash red envelope activity report shows that from February 1 to February 17, there were over 3.6 billion lottery draws and users completed AI tasks over 1 billion times, with 49% of users coming from third and fourth-tier cities [2] - Netflix's co-CEO Ted Sarandos stated that generative AI will assist creators by speeding up production times rather than harming job prospects, emphasizing that AI tools can enhance storytelling capabilities [2] - Alphabet announced that the Google I/O developer conference will be held from May 19 to 20, where updates on the Gemini large model and other AI products are expected, along with the potential launch of smart glasses [2] Group 2 - Moore Threads announced full adaptation of Alibaba's latest large model Qwen3.5 on its flagship AI training and inference GPU MTT S5000, showcasing the maturity of the MUSA ecosystem [3] - The adaptation process validated two core capabilities of the MUSA ecosystem: native MUSA C support for kernel development and deep compatibility with Triton-MUSA for high-performance operator writing [3]

Core Insights - Moore Threads has successfully adapted its flagship AI training and inference GPU MTT S5000 to Alibaba's latest large model Qwen3.5, showcasing the maturity and completeness of the MUSA ecosystem [1] Group 1: MUSA Ecosystem - The adaptation demonstrates two core capabilities of the MUSA ecosystem: native MUSA C support allows developers to directly use MUSA C for kernel development, significantly lowering the migration barrier from the CUDA ecosystem [1] - Deep compatibility with Triton-MUSA enables developers to write high-performance operators using familiar Triton syntax, which can seamlessly run on Moore Threads' full-featured GPUs through the Triton-MUSA backend [1]

Group 1 - Moore Threads has successfully achieved Day-0 rapid adaptation of the MiniMax M2.5 model on its flagship AI training and inference GPU, the MTT S5000 [2] - Elon Musk predicts that by the end of 2026, AI will bypass traditional programming and directly generate binary programs, claiming that AI-generated binaries will outperform any compiler-generated results [2] - Spotify's CEO stated that some of the company's top developers have not written any code this year, as all code has been generated by AI, indicating a significant shift in the development process [2] Group 2 - Elon Musk announced that the Grok 4.20 version will be officially released next week, featuring significant improvements compared to version 4.1 [2]