
A | 2026世界人工智能大会(WAIC)期间,摩尔线程披露,一款236B参数的MoE模型基于国产万卡级集群训练,一款具身智能大脑模型,基于国产千卡集群训练,此外,北京大学EvoPhys团队首个全球5D世界模型EvoPhys-World的全栈原生训练,全程由MUSA软件栈提供底层支撑。 Tamil cinema is gearing up for a major nostalgia wave as the cult Tamil hit film Padayappa storms back into theatres on 12 December, perfectly timed with Rajinikanth’s 75th birthday and his 50-year milestone in cinema. For fans, this is not just a rerelease but an invitation to revisit a cultural moment that helped define superstardom, fan hysteria, and the sheer spectacle of big screen entertainment.In a newly surfaced behind the scenes video, Rajinikanth looks back on the making of the KS Ravikumar hit with characteristic candour. He recalls being astonished by the sheer volume of women who turned up at cinemas during its original run, something he admits he had never witnessed in his decades long career. For a star known for pan demographic appeal, that detail captures the special emotional tether Padayappa created on release.Rajinikanth Says Padayappa 2 Is Under Serious ConsiderationWhat has amplified fan frenzy even more is Rajinikanth’s confirmation that discussions for Padayappa 2 have begun. Having already been part of sequels like Jailer 2 and 2.0, he said he strongly felt the world of Padayappa still had unexplored dimensions.He added that the proposed sequel would shift its gaze entirely to Neelambari. Ramya Krishnan’s iconic antagonist continues to be one of Tamil cinema’s most celebrated characters, remembered for her intensity, elegance, and fierce emotional complexity. Built around her vow to avenge Padayappa even through another birth, the potential storyline has already ignited rampant speculation among fans.Rajinikanth shared that the creative team is currently shaping the narrative and that the new chapter, if finalised, will retain the festive scale and emotional fireworks that made the first film a landmark in his career. A Film Made for the Big Screen ExperienceThe superstar also spoke about the sacredness of the theatrical experience while making Padayappa. He revealed that he consciously resisted lucrative digital offers because he believed the film deserved to be experienced only in cinemas. Except for selling satellite rights, he kept the film away from early home viewing platforms, insisting that its visual energy and narrative punch were designed for the big screen.Directed by KS Ravikumar, the 1999 blockbuster brought together an ensemble that still feels timeless. Sivaji Ganesan delivered regal gravitas as Padayappa’s father, while Soundarya offered warmth and restraint opposite Rajinikanth. Radha Ravi, Nassar, Manivannan, and Abbas rounded out a cast that propelled the story’s blend of conflict, romance, and redemption. View this post on Instagram A post shared by Soundarya Rajinikanth (@soundaryaarajinikant) AR Rahman’s soundtrack was a sensation on release and remains one of his enduring commercial albums. Tracks like En Peru Padayappa and Minsara Poove continue to dominate nostalgic playlists, testimony to the musical legacy that accompanied the film’s theatrical impact.More than 25 years later, Padayappa still represents everything that makes mass cinema a communal celebration. The emotion, the punch lines, the star power, and the myth making around Rajinikanth have aged into legend. Its rerelease arrives at a symbolic moment in his career, allowing audiences to relive the energy that once turned theatres into festival grounds.And if a Neelambari centred sequel does take shape, fans may soon witness a fresh chapter in one of Tamil cinema’s most iconic rivalries.Also Read: Tamil Cinema Bids Farewell to AVM Saravanan, A Producer Who Shaped Generations。 其中,MoE-236B模型基于25T+语料,在摩尔线程国产万卡级集群上从零起步完成预训练到长窗口训练的全流程,有效训练时长占比超过90%。

B | “高精度”是摩尔线程反复强调的能力,摩尔线程高级副总裁董龙飞提到,它指的不是单次计算的数值精度,而是模型在不同GPU平台间迁移训练时的结果对齐能力,即相同模型架构、训练数据、超参数下,能否获得稳定、可比的训练结果。 摩尔线程给出的具体数据是,在DeepSeek一类MoE模型上,与国际主流GPU做对比训练,前3万步的平均相对误差控制在万分之六以内。摩尔线程还以500B至5T不等规模的数据做过多轮实验,称基于相同卡数、模型配置和训练流程,摩尔线程平台训练所得模型在下游Benchmark评测中的得分与参考平台基本一致,部分指标更高。 摩尔线程与北京大学EvoPhys团队合作训练的5D世界模型EvoPhys-World,基于MTT S5000完成全栈原生训练,在WorldScore“世界生成”维度连续37天排名第一,该项目获得摩尔线程“灯塔计划”支持,这是国内第一次由中国科学家在中国算力设备上完成的原创世界模型工作,也是“国芯训国模”的一个分水岭。 此外,北京智源研究院的RoboBrain 2.5具身大脑模型,基于FlagOS-Robo框架,依托摩尔线程MTT S5000千卡智算集群,双方成功完成RoboBrain 2.5的全流程训练。首次验证了国产算力集群在具身智能大模型训练中的高可用性与效率。 摩尔线程创始人、董事长兼首席执行官张建中将AI产业按功能拆分为三类“工厂”:模型训练工厂、词元生产工厂、智能体生产工厂。三者共用同一套全功能GPU架构,这也是摩尔线程与专用芯片路线(ASIC)厂商的核心分野。

C | 词元生产工厂方向,摩尔线程重点展示了基于MTT S5000的PD分离异构推理方案,将算力需求大的Prefill(预填充)计算放在S5000上,将带宽需求高的Decode(生成)计算放在国际主流GPU上,两者异构分池部署。 张建中在演讲中给出的数据是,异构组合后整体吞吐量较原有方案提升近2倍,并给出算式,“3台S5000+2台国际主流GPU≈9台国际主流GPU”。董龙飞补充表示,这类异构方案的性能比任何同构方案都要高50%以上,其中很大一部分收益来自KV Cache的调度优化,在类似OpenClaw这样反复调用本机命令行工具的场景下,缓存命中率可达90%,相当于节省90%的算力。 据悉,过去一年,国产大模型发布后到完成硬件适配的周期已从“数月”压缩到“发布即适配”(Day 0适配),部分大模型厂商现在会在模型规划部署阶段就与摩尔线程同步“共研”,而不是等模型发布后才启动适配工作。 在具身智能方面,早期VLA/VLM类模型参数在5B至14B之间,主要部署于端侧;当前的世界模型路线不再依赖语言输入,直接处理物理世界,参数跨度更大,实现高质量具身控制大致需要10亿参数级别模型在端侧运行,同时配合千亿级(1000B)及以上模型在云端协同。摩尔线程正基于自有万卡级集群持续进行万亿参数模型的训练验证,但未公布具体进展和时间表。 提及算力供应话题,董龙飞表示,中国的算力建设是面向未来十至二十年的基础设施投入,短期供不应求属于正常现象,以铁路、高速公路等基建类比,中国GPU或算力企业远未发展到“供过于求”的地步。

D | (本文作者 | 张帅,编辑 | 杨林) 更多精彩内容,关注钛媒体微信号(ID:taimeiti),或者下载钛媒体App。
Current article:http://lindunfokangliucishirun.shop/rogzsw/20260826/1281.html
Published on:00:59:30
三星发布新一代折叠屏智能手机,消费电子创新周期再度开启
微软的新专利:表面拨号2或触摸支持
Bangladesh's Gaibandha bypolls stopped: Poll body
新研究表明北大西洋或正在接近一个新的“临界点”
华硕公布 TUF Gaming K4 Magnetic 磁轴键盘:天选 TX75 兄弟型号
市农业农村局:全速“冲刺跑” 以提质增效助推农业高质量发展