Key Takeaways TEVA's new drugs Austedo, Ajovy, and Uzedy are driving strong branded sales growth across markets. TEVA plans to double biosimilar sales by 2027, with multiple U.S. and EU launches lined up. Cost cuts, pipeline progress, and debt reduction have lifted investor sentiment.Teva Pharmaceutical Industries Limited’s (TEVA) shares have risen 21.6% in three months.  Over the past few quarters, Teva has successfully launched several biosimilars and other high-value, complex generics, with many more i ...

Larimar Therapeutics Nomlabofusp Regulatory Update June 2025 1 pro Forward-Looking Statements This presentation contains forward-looking statements that are based on the beliefs and assumptions of Larimar Therapeutics, Inc. ( "Company") and on information currently available to management. All statements contained in this presentation other than statements of historical fact are forward-looking statements, including but not limited to Larimar's ability to develop and commercialize nomlabofusp (CTI-1601) and ...

BALA CYNWYD, Pa., June 20, 2025 (GLOBE NEWSWIRE) -- Larimar Therapeutics, Inc. (Larimar) (Nasdaq: LRMR), a clinical-stage biotechnology company focused on developing treatments for complex rare diseases, today announced that the Company will host a conference call and webcast to discuss regulatory updates for the Company’s nomlabofusp clinical development program for the treatment of Friedreich’s Ataxia on Monday, June 23, 2025 at 8:00 am EDT. Conference Call and Webcast DetailsTo access the webcast on Mond ...

6月6日，"鸿日山海跑"——鸿日汽车集团30周年庆活动在山东五莲、安徽金寨两地同时开跑。"以'山海跑'之名，致敬过往30年的征程，更以奔跑之 姿，开启下一个30年的华章。"正如鸿日汽车集团董事长张建军所言，经过30年的"长跑"，鸿日汽车集团已经昂首跑进微型电动汽车的新发展阶段。 围绕市场需求 提升产品品质 鸿日汽车推出的行业首个智能化架构平台——"Hi平台"，具有"便捷"、"节能"、"舒适"、"安全"四大特点，以此平台为基础开发的各款车型，在动 力、智能化、舒适性等方面都拥有不错的表现。按照规划，今年鸿日汽车集团将进一步提升产品的智能化水平，根据市场需求推出"数智化透视底 盘"、"智能语音"、"遥控泊车"、"冷暖空调"、"加热杯托"等技术，让新车型全面满足消费者对智能化的需求。 在行业发展过程中，很多低速电动车并没有如新能源乘用车一样执行严格的试验验证标准。为了更好的提升、验证自身产品的性能，鸿日汽车集团主 动对自身产品展开各类严苛的试验验证，并邀请第三方机构进行公开测评，向消费者展示产品的真实性能。近年来，各类媒体驾驶鸿日汽车集团的产品驶 上高原、开展冬测等试验频繁冲上热搜，让消费者近距离了解到了鸿日汽车的 ...

Group 1 - The 30th Guangzhou International Lighting Exhibition (GILE) took place from June 9 to 12, 2025, featuring over 3,000 exhibitors and focusing on the theme "AI + 3D Printing: Pioneering a New Era of Lighting" [1] - YD Illumination showcased a 200+ square-meter double-decker exhibition hall, highlighting its latest innovations in lighting and AI-powered 3D printing [1][5] - The exhibition included immersive simulations demonstrating the durability of YD Illumination's IP68 lighting products in various harsh environments [3] Group 2 - A notable exhibit was the IP68 Flexible Strip Light Series, which won the Aladdin Golden Lamp Award for National Excellence, alongside the Mini Master-sub controller [3] - An interactive "Mermaid" installation was created using AI-generated 3D modeling and precision 3D printing, allowing real-time human-light interactions [4] - The exhibition featured an AI + 3D Printing Experience Zone where visitors could generate 3D models from text descriptions or images, followed by live demonstrations of the printing process [4] Group 3 - YD Illumination is committed to technological innovation and aims to deliver groundbreaking solutions while exploring the possibilities of light [5]

在大语言模型蓬勃发展的背景下，Transformer 架构依然是不可替代的核心组件。尽管其自注意力机制存在计算复杂度为二次方的问题，成为众多研究试图突破的 重点，但 Transformer 在推理时灵活建模长距离上下文的能力，使得许多线性复杂度的替代方案（如 RNN、Linear Attention、SSM 等）难以真正取代它的地位。 尤其是在大语言模型广泛采用 decoder-only 架构之后，自注意力机制的重要性进一步凸显。然而，这种机制也带来新的挑战：推理过程中每一步都需要访问 Key- Value（KV）缓存，该缓存的大小随着生成序列长度线性增长，逐渐成为影响推理效率的关键瓶颈。随着模型参数维度不断扩大，KV 缓存所需的显存和带宽开销 显著上升，限制了模型的推理长度与可支持的 batch size。 值得一提的是，近期由 DeepSeek 团队提出的 MLA 机制，通过在隐空间维度对 KV 缓存进行压缩，显著提升了推理效率，推动了大模型在低资源场景下的高效部 署。但随着生成序列的持续增长，时间维度的冗余信息也逐渐暴露，压缩其所带来的潜力亟待挖掘。然而，如何在保持性能的前提下压缩时间维度，一直受到增 ...

Summary of Alvotech (ALVO) Conference Call Company Overview - Alvotech is a leading pure play biosimilars company focused on providing access to low-cost biologic drugs globally, with a strong emphasis on high quality [3][4] - The company has established a significant presence in the biosimilars market over the past ten years, with a total addressable market for biologics estimated at $200 billion [4] Core Business Strategy - Alvotech employs a partnership approach, collaborating with strong local partners in various countries to enhance market reach [4][16] - The company has invested heavily in R&D and manufacturing capabilities, integrating both functions to streamline operations [14][15] - Alvotech has a multiproduct portfolio strategy, with a focus on biosimilars, which are projected to grow significantly as patents for biologics expire [7][10] Market Opportunity - Approximately 120 biologics are expected to lose patent protection in the next ten years, creating a substantial opportunity for biosimilar development [7] - The biosimilar market has been growing at around 20% over the last three years, particularly in immunology and oncology [11] Financial Performance and Projections - Alvotech reported revenues of approximately $275 million last year, with expectations to grow to between $600 million and $700 million this year [28] - EBITDA is projected to increase from $108 million to between $200 million and $280 million this year, indicating strong operating leverage [28] - By 2028, Alvotech aims to achieve revenues exceeding $1.5 billion, with an EBITDA margin of 40-45% [29][30] Product Pipeline and Launches - Alvotech has launched two biosimilars in the U.S. market: Humira (ABTO2) and Stelara (AVTO4) [19][20] - The company anticipates three new approvals and launches by the end of the year, with a total addressable market of around $5 billion [21] - Future launches include biosimilars for Entyvio, high-dose Eylea, and others, expected to contribute significantly to revenue targets [22][33] Strategic Partnerships - Alvotech has established 19 global partnerships, which provide steady cash flow and support R&D funding [16] - The company emphasizes deepening existing partnerships to leverage market opportunities effectively [42][43] Risk Management - Alvotech does not foresee significant impacts from U.S. tariffs due to favorable tariff rates and contract structures [25] - The company is confident in its ability to navigate patent discussions and maintain a competitive edge in the biosimilar market [37][41] Conclusion - Alvotech is positioned as a strong player in the biosimilars market with a robust growth strategy, significant market opportunities, and a comprehensive product pipeline [35]

一水 发自 凹非寺 量子位 | 公众号 QbitAI 曾撼动Transformer统治地位的Mamba作者之一 Tri Dao ，刚刚带来新作—— 提出两种专为推理"量身定制"的注意力机制。 在保持模型性能不变的情况下，将解码速度和吞吐量最高提升2倍，大大优化了模型的长上下文推理能力。 这项研究的三位作者均来自普林斯顿大学，论文主要有两大贡献： 其一，提出Grouped-Tied Attention（GTA） ，与已集成到LLaMA 3的注意力机制GQA质量相当，但KV缓存用量减少约50%。 其二，提出Grouped Latent Attention（GLA） ，与DeepSeek所使用的注意力机制MLA质量匹配，但解码速度更快，某些情况下比 FlashMLA快2倍。 按照作者之一Ted Zadouri的总结： GTA是GQA的有效替代品，而GLA是MLA的实用替代品。 一言以蔽之，通过优化注意力机制的 内存使用 和 计算逻辑 ，在不牺牲模型生成质量的前提下，可显著提升大语言模型的推理效率和硬件资 源利用率，尤其在长上下文场景中优势更为突出。 相关论文公布后，一众研究者也赶来祝贺~ | Sara Hooker ...

HUAWEI X HUXIU AMLA × 融合算子 × SMTurbo 让大模型推理速度与能效双重革命 没有人不在期待大模型能够成为下一个电动车，作为代表中国的新兴产业，在世界范围内掀 起狂澜。 然而主流的MoE架构大模型，却苦于其结构上的"先天不足"：巨大的硬件成本与多重拖累效 率的环节，使得中国企业在这场芯片堆砌与效率挖掘的苦径上难以提速。 作为智能基础设施提供商，华为在这场战役中另辟蹊径，利用其在数学算法和工程领域的深 厚积累，为DeepSeek显著提升了效率及用户体验。 山就在那里，但中国企业找到了不一样的登顶之路。 近期，虎嗅将打造《华为技术披露集》系列内容，全面揭秘超大规模MoE模型推理部署技 术，通过一连串的技术报告，首次全面披露技术细节。 希望本系列内容能为业界起到参考价值，也希望更多人能与华为一起，共同打造长期持续的 开放协作生态环境，让昇腾生态在中国茁壮成长。 《华为技术披露集》系列 VOL.5 ：昇腾亲和 它们不仅是模型的效率引擎，更是硬件性能的放大器 —— 通过标准化设计、硬件深度适配与 复用机制，让芯片处理海量数据时如虎添翼。 而昇腾此次开源的三大技术，正是算子优化的 "终极形态" ...

" 算力利用率突破 70%、跨卡延迟低至亚微秒级，华为算子技术 如何重新定义硬件性能？ " 作者丨李希 AMLA × 融合算子 × SMTurbo —— 让大模型推理速度与能效双重革命 基于昇腾算力，华为团队本次发布了三项重要的硬件亲和算子技术研究： 针对Decode阶段的MLA 计算，华为团队提出了AMLA（Ascend MLA）算子，通过数学等价变化和硬件 亲和的深度优化， 释放昇腾芯片澎湃算力 。具体而言，通过 对浮点数二进制编码的重解析 ，把复杂的乘 法运算变成简单的加法操作，AMLA实现了基于 存内计算 的变量更新，充分利用算力的同时减少数据搬 运；结合一系列基于昇腾硬件的计算流程及流水优化手段，进一步提升算子的整体性能。当前AMLA算法 的Attention 算子充分发挥昇腾硬件的计算能力，平均算力利用率达到55%，最高可达71%，优于 FlashMLA公开的结果。 大模型推理的 "乐高积木"：算子为何如此重要？ 算子是 AI 大模型执行计算的 "原子级工具"，如同乐高积木中的基础模块，负责从加减乘除到特征提取的 一切核心操作。它们不仅是模型的效率引擎，更是硬件性能的放大器 —— 通过标准化设计、 ...