大模型前沿速递

Enterprise AI deployment accelerating with platform engineering, harness engineering, open-weight models. DeepSeek Harness open source. Dynatrace acquires Arize. AI funding surpasses semiconductors. Nvidia/Meta local models. Google abandons frontier race. DeepSeek price hike. Open ecosystem shift. Australian adoption. Qwen3.8-27B, Max open-sourced. Alibaba skill-up. Nvidia financing platform. Post-training data supply chain. RAG cost optimization. Anthropic premium pricing. Stripe acquires OpenRouter for $8B. OpenAI DeployCo. A2A protocol. Agentic AI production data. F5 AI Gateway reduces token spend 60%. Snowflake Cortex AI integrates DeepSeek/GLM with dynamic routing. Enterprises ditch frontier models for data sovereignty. DeepSeek Harness plugin ecosystem expands. Nvidia releases NeMo Switchyard. Open-source models shift to revenue-sharing. Meituan open-sources internal AI methodology. F5 AI Gateway integrates model gateway, MCP gateway, and AI guardrails. MixT compression method from CAS achieves near half parameter reduction. End-side AI OS thinking article. Ling 3.0 Flash open-weight model surpasses GPT-5 in intelligence and cost. Stripe acquires OpenRouter for $8B, validating multi-model routing and token financialization. Liquid AI releases DSpark draft models for 3.18x faster decoding. Chinese open-source ecosystem expands with DeepSeek Harness, Xiaohongshu 280B agent, MiniMax music model. Automotive AI competition shifts to chain efficiency. AT&T adopts model routers and open-source models (Nemotron, Llama, Gemma) to cut costs, explicitly avoiding Chinese models due to risk. Anthropic's 30-day data retention mandate draws enterprise pushback. 钛动科技 and 智合AI raise funding for AI marketing and legal verticals. Ramp launches its own model router, Router, signaling model routing as competitive infrastructure layer. ExcelScale launches in Malaysia with unified API for 100+ models. Falling AI costs chart shows dramatic price drops, reinforcing 'good enough' model selection logic. Healthcare CIOs struggle with Epic's tokenomics. Alibaba earnings show AI payback: Cloud revenue up 45%, AI products 35% of external cloud, $10B quarterly CapEx. Nvidia partners with Apollo/BlackRock to finance AI infrastructure, mobilizing $500B. Microsoft releases Agent Lightning v1.0, achieving 14.6-point SWE-bench gain on Qwen3.5-9B. DeepSeek weekend discount pricing and V4 Pro agent capabilities further drive cost optimization. Gemma downloads surpass 1B. AI discovers novel cancer mechanism via C2S-Scale model. Huawei and iFlytek expand to Brazil with WAICO framework. US AI companies lower prices in response to Chinese open-source pressure. DeepSeek V4-Flash multimodal with agent improvements. Nvidia research shows agent harness design (AVO system) can matter more than model choice. Ex-DeepMind team's Faraday agent using 27B Qwen 3.6 outperforms GPT-5.5 on paper reproduction. Kimi K3 open day reveals detailed technical report and three Infra technologies open-sourced. PlayWorld benchmark for world models. Wuhan compute platform upgrades to tokenized trading and one-stop model access. Mira Murati's new company open-sources Inkling Small (276B/12B MoE, million-token context). GPT-5.6 price cut 80%. Kimi K3.1 announced for August release. Micron to invest $10B in memory R&D, signaling memory bandwidth as next AI bottleneck; partnership with Anthropic. Dyna Robotics DYNA-2 World-Action Model shows video scaling for embodied AI. Poolside Laguna S 2.1 open-weight coding model. 字节发布豆包工作智能体，与飞书深度打通，强调企业上下文理解，标志AI应用大规模商业化落地。汤森路透用4000万美元在开源底座上训练法律大模型Thomson，验证数据护城河+专家监督+模型工厂流水线路径。保险业AI垂类大模型案例（泰康、太平再保险）展示开源浪潮下中小险企技术平权。RoboColiseum标准化仿真平台加速具身AI评估。小米发布玄戒O100AI芯片，小鹏机器人融资9亿美元，智元机器人夺冠等产业动态。 Samsung uses Anthropic Claude for chip design, reporting both productivity gains and serious errors. OpenAI self-chip Jalapeno beats Nvidia GB300. NVIDIA Vera Rubin with DeepSeek V4 Pro achieves 30x agent performance. AI infrastructure enters debt financing cycle; credit risk not yet elevated. New articles: DeepSeek Harness curated list (reference); agentic AI architecture primer; 2026 World Robot Conference roundtable on world model data challenges. Recent article on AI price cuts accelerating application adoption (Luna price cut 80% led to 14x call volume). Aerospace vertical model successfully tested on-orbit (94-96% accuracy). Enterprise agent platform case studies (瓴羊AgentOne) show practical deployment. New: IBM Granite 4.2 models (3B/8B/30B, 128K ctx, agentic RL) for local deployment. Qwen3.8-Flash with 1/9 training cost. Enterprise AI SRE strategies and observability challenges. Day-One Architecture article emphasizes decoupling models and composable infrastructure. Small specialized models eating AI stack: most inference done by small models, frontier only for hard reasoning. Enterprise data governance becomes agent bottleneck. 王云鹤创业基元律动押注Routing Harness，日Token调用5000亿，DRACO评测1/3成本超Fable5。Veeda AI获9000万美元种子轮打造机器人世界模型。蜂巢互联12亿元融资加码物理AI仿真。Perplexity估值300亿美元向Agent转型。