## 🔍 Key Insights **Kimi K2.5** has emerged as a new global benchmark for infrastructure-grade AI deployment—Cloudflare has integrated the model into its core production systems, achieving a **77% reduction in operational costs**, while simultaneously enabling AI Agents and automated code review [19]. Concurrently, **IBM Granite 4.0 3B Vision**, built on a modular **DeepStack architecture** and trained on the proprietary **ChartNet dataset**, overcomes longstanding bottlenecks in enterprise document understanding—underscoring the accelerating momentum behind lightweight, multimodal AI deployment [0]. ## 🚀 Major Updates - **IBM launches Granite 4.0 3B Vision**, a compact multimodal model optimized for table extraction and chart comprehension, powered by the DeepStack architecture [0] - **Cloudflare fully integrates Kimi K2.5** [19]: deployed for AI Agents and code review—cutting costs by 77% and serving ~20% of all websites globally - **Runway officially launches the Runway Fund and the Builders Program** [4][5]: dual-track initiative supporting early-stage startups in AI-powered media creation and world simulation—from seed to Series C - **Qdrant releases Agent Skills**, a vector retrieval augmentation toolkit that significantly improves retrieval success rate, latency, and token efficiency in agent systems [17] - **Claude Code source code accidentally leaked** due to an npm map configuration error [9]: prompting deep community reflection on supply-chain security for generative AI tools - **Bittensor completes training of a 72B-parameter LLM on its permissionless network** [11]: marking a milestone where decentralized large-model training becomes engineering-feasible - **Stanford & MIT introduce Meta-Harness**, an automated framework for building AI agent shells [13]: dramatically boosting engineering efficiency for NLAHs (Natural Language Agent Shells) - **Zhipu AI releases its first post-IPO financial report** [21]: full-year revenue reached **¥724 million**, leading domestic large-model commercialization—and validating the high-value closed-loop potential of the MaaS (Model-as-a-Service) model ## 🔗 Sources [0] Granite 4.0 3B Vision: A Compact Multimodal Intelligence Engine for Enterprise Documents — https://www.bestblogs.dev/article/75d40607 [4] Runway Launches Its Venture Fund — https://www.bestblogs.dev/status/2038984561132990836 [5] Runway Announces the "Runway Builders" Program — https://www.bestblogs.dev/status/2038983401982619653 [9] Claude Code Source Code Leaked Due to npm Map Configuration Error — https://www.bestblogs.dev/status/2038978960466272398 [11] Bittensor's 72B-Parameter LLM Training Milestone — https://www.bestblogs.dev/status/2038973633280803206 [13] Meta-Harness: Automating AI Agent Shell Engineering — https://www.bestblogs.dev/status/2038967842075500870 [17] Qdrant Releases Agent Skills to Enhance Vector Search Capabilities — https://www.bestblogs.dev/status/2038964537370738811 [19] Cloudflare, Serving ~20% of All Websites Globally — https://www.bestblogs.dev/status/2038962222222222222