【20251220AI日报】上海市规划资源局与商汤大装置联合打造，“云宇星空大模型（专业版）”正式发布

本文字数：约 3600 字，预计阅读时间：12 分钟上海市规划资源局与商汤大装置联合打造，“云宇星空大模型（专业版）”正式发布近日，全国规划资源领域首个基础大模型“云宇星空大模型（专业版）”在上海正式发布。该模型由上海市规划资源局联合商汤大装置共同研发。云宇星空大模型（专业版）基于海量高质量的“文本-图像-空间”多模态行业语料，进行了城市空间治理领域全链路认知与决策能力的突破性构建，形成了集“时空理解生成、知识整合检索、模型智能调度”于一体的技术底座。具备“问不倒、能调图、会统计、能识图、会报告”五大核心专业能力，成为上海规划和自然资源行业专业人员处理各类专业问题的“AI伙伴”。该大模型打造了“基座大模型+六大垂类模型”的“1+6”模型体系，兼顾通用能力与行业深度，支撑复杂空间治理场景下的专业智能需求。目前，该模型已经覆盖规划资源、政府治理、社会共创三大领域的十大核心场景，可全方位赋能规划和自然资源行业的专业工作。云宇星空大模型（专业版）在规划资源领域，能够服务于城市总体规划、土地精准供给、产业用地增效等场景，实现规划编制智能化、资源调配精准化，助力国土空间治理现代化；在政府治理领域，

今日新鲜事 · 2025-12-19

【20251219AI日报】Palona goes vertical, launching Vision, Workflow features: 4 key lessons for AI builders

本文字数：约 4000 字，预计阅读时间：20 分钟Palona goes vertical, launching Vision, Workflow features: 4 key lessons for AI buildersPalona AI, a Palo Alto-based startup, has recently launched Palona Vision and Palona Workflow, which transform the company's multimodal agent suite into a real-time operating system for restaurant operations. Initially, the company aimed to build emotionally intelligent sales agents for broad direct-to-consumer enterprises. However, the shift to a restaurant and hospitality focus dem

今日新鲜事 · 2025-12-18

【20251218AI日报】Gemini 3 Flash arrives with reduced costs and latency — a powerful combo for enterprises

本文字数：约 3825 字，预计阅读时间：15 分钟Gemini 3 Flash arrives with reduced costs and latency — a powerful combo for enterprisesGoogle has introduced Gemini 3 Flash, a large language model (LLM) that offers a significant reduction in cost and latency compared to its predecessor, Gemini 3 Pro. This new model is designed to be more accessible and efficient for enterprise use. The introduction of Gemini 3 Flash is a strategic move to make high-quality AI models more affordable and faster, enabling companies to d

今日新鲜事 · 2025-12-17

【20251217AI日报】Zencoder drops Zenflow, a free AI orchestration tool that pits Claude against OpenAI’s models to catch coding errors

本文字数：约 3800 字，预计阅读时间：18 分钟Zencoder drops Zenflow, a free AI orchestration tool that pits Claude against OpenAI’s models to catch coding errorsZencoder, a Silicon Valley startup that develops AI-powered coding agents, recently released Zenflow, a free desktop application designed to transform how software engineers interact with artificial intelligence. This innovative tool aims to move the industry beyond the current trend of "vibe coding" by introducing a more disciplined and verifiable approac

今日新鲜事 · 2025-12-16

【20251216AI日报】Korean AI startup Motif reveals 4 big lessons for training enterprise LLMs

本文字数：约 3300 字，预计阅读时间：15 分钟Korean AI startup Motif reveals 4 big lessons for training enterprise LLMsMotif Technologies，一家韩国的AI初创公司，最近发布了一个名为Motif-2-12.7B-Reasoning的模型，并分享了他们在训练企业LLM时的四大重要经验。这些经验对希望构建或微调自己模型的企业团队有着极大的参考价值。第一，推理能力的提升来自于数据分布而非模型大小。Motif发现，合成推理数据只有在与目标模型的推理风格匹配时才有效。这意味着企业内部团队不应简单地复制外部数据集，而应确保合成数据符合推理阶段所需的格式、冗长程度和步骤细节。第二，长上下文训练首先是一个基础设施问题。Motif在Nvidia H100级硬件上实现了64K上下文训练，这依赖于混合并行性、仔细的分片策略以及激进的激活检查点技术。这对于希望构建具有长上下文能力模型的企业来说是一个重要的提醒：长上下文能力需要在训练堆栈的早期设计中考虑。第三，强化学习微调需要数据过滤和重用。Motif的强化学习微调

今日新鲜事 · 2025-12-15
2022 - 2026 康叔的AI全栈工坊. All Rights Reserved.
Theme Jasmine by Kent Liao