大模型前沿速递

Writer's AI harness reduces costs. Anthropic bets $1.5B on deployment. Core battleground shifting to platform engineering. Box report shows 83% running agents but only 36% connected to trusted content. Harness engineering confirmed as key differentiator: AI Agent Harnesses Can Change Red Teaming Results (Kimi-K2.6 attack success jumped from 1% to 24% by switching harness). 实在Agent tops OSWorld with 90.2% success rate. DeepSeek retrained V4-Flash beats flagship Pro on agent benchmarks. OpenAI's Presence signals vertical integration into agent infrastructure. Microsoft previews MAI-Realtime bidirectional voice model. GPU utilization optimization techniques gain traction. Evaluation shifting from external benchmarks to internal product loops. Chinese AI ecosystem: 上海AI生态持续繁荣. MiniMax H3 open-source adds to ecosystem; LMArena analysis highlights evaluation market. Token pricing analysis reveals cost structure opacity and standardization push; 信通院攀登计划基线 referenced. Marvell shares surge 8%+ on AI memory/storage solutions (CXL/PCIe Gen6). 面壁智能 ForgeStencil dual-agent industrial software optimization (hypre 3.86x, FDTD 2.47x). 明略科技 enterprise agent deployment case study. Marvell advances AI memory infrastructure with CXL and optical shared memory; CIOs advised on model churn management with abstraction layers; Inkling-Small (276B MoE, 120B active) released at 1/4 size and 1/3 cost, supporting single-GPU deployment; AI agents rebuild 16,000-line programs from scratch (Claude Fable 5 64% solve rate, $251 cost). Thomson Reuters' homegrown legal AI model claims to rival frontier labs, highlighting domain-specific model narrowing the gap. Comprehensive analysis of Chinese AI Agent landscape confirms product stratification and ecosystem maturity. AI agents challenge identity governance, requiring runtime controls and intent deviation detection. HappyRobot raises $150M on enterprise agent architecture (multi-model orchestration, 31% failure from resolution breakdowns). Financial infrastructure vendors emerge as hidden winners in AI boom. Alibaba's Qwen3.8-Max integrated with DingTalk for enterprise office AI, competing with Tencent and ByteDance. Snyk report: enterprise AI footprint 3x larger than model lists, 46.9% adopting agentic architectures. European established tech firms (SAP, Capgemini) emerge as unexpected AI winners, highlighting value accrual to integrators. New: Peking University and YuanKong AI open-source OpenAI4S, a research agent replicating Claude Science, MIT license, zero-dependency, 30+ research skills. Liquid AI releases LFM2.5-2.6B end-side model, surpassing larger models on agent tasks. ByteDance releases SeedRealtime, a unified real-time multimodal model. Claude Opus 5 with effort knob offers half-cost near-Fable5 performance. Shanghai AI Lab's Mobius architecture addresses Transformer limitations with knowledge-reasoning separation, 4x inference efficiency. New: Hark previews browser use agent with $700M funding, claims speed/cost advantages over GPT-5.5 and Opus 4.8 using action-prediction approach. Jeff Dean and Sanjay Ghemawat leave Google to co-found Discovery Loop, aiming to automate scientific discovery, signaling major talent migration to AI infrastructure. New data: China manufacturing AI Agent market reaches 135B yuan, with 实在Agent, Alibaba Cloud, Huawei Cloud leading. End-side AI in phones and cars accelerating. New: Nvidia showcases Cosmos 3 and Omniverse for open physical AI development, adding physical AI dimension. New: AMD acquires Taalas to hardwire AI models into silicon, achieving 73x speedup at 1/10 power, reshaping inference hardware landscape. New: Model routing panel from NVIDIA, Cognition, OpenRouter offers practical cost optimization strategies. New: NVIDIA releases NOOA framework for agent development, collapsing complexity into single Python class. New: Microsoft Build 2026 reveals 7 self-developed MAI models, including MAI-Thinking-1 (350B params, AIME 25 97%) and MAI-Code-1-Flash (5B params, SWE-Bench Pro 51%), marking a shift from OpenAI dependency to self-reliance. New: ZhikunCode, an open-source AI coding tool alternative to Claude Code/Cursor, released on GitHub, supporting multiple Chinese models with 56% SWE-bench resolve rate. New: Arena AI CEO confirms enterprises don't know which models to trust, driving demand for model routing and abstraction layers. New: Intel releases OpenVINO 2026.3 with MoE disk offloading for 30B models on 16GB, enabling edge deployment. New: DeepSeek API price increase due to compute costs, affecting enterprise cost calculations. New: Apple selects Alibaba Qwen for China iPhone AI, validating commercial traction. New: Brokerage firms (券商) accelerate AI agent deployment with 32 registered products, signaling financial sector adoption.