IBM Research released Granite Timeseries Ensemble R1 on Hugging Face, providing a zero-shot ensemble forecasting solution based on four pre-trained Granite time series models (PatchTST-FM-r1/r2, FlowState-r1.1, TTM-r3). It supports two default aggregation methods: Probabilistic Ensemble (equal-weight probability pooling) and IQR Ensemble (adaptive interval weighting), requiring no fine-tuning or weight fitting. Validated on 97 GIFT-Eval configurations, it demonstrates superior combined MASE and CRPS/MWQL performance compared to individual member models. It can be installed via pip through the tsfm_public library and loaded with pre-configured recipes, outputting point forecasts and quantile forecasts from 0.1 to 0.9.
FluxDown author clow released version v0.6.0, introducing a universal plugin system for desktop and web to its open-source multi-protocol downloader (AGPL-3.0, Rust engine, supporting HTTP/BT/HLS/DASH/FTP/eD2K).
Aw-1 is a zero-dependency Python security circuit breaker that intercepts AI agent tool calls within sub-millisecond loops via Abstract Syntax Tree (AST) decompilation, policy-as-code, rolling rate monitoring, and HMAC-signed execution tokens; version v2.1.0 achieved a throughput of 8485.0 dispatches/sec and P99 latency of 0.2263 ms in a stress test with 10,000 concurrent asynchronous dispatches; it supports local CLI zero-install trials, @guard decorator integration, LangGraph adaptation, and production-grade HTTP gateway services; open-sourced on GitHub under the Apache 2.0 license.
The iFlytek Spark X2.6 end-cloud integrated large model was released at the iFlytek 1024 Developer Festival, adopting a 293B-A30B parameter MoE architecture with a context length of up to 1M. Trained on fully domestic computing power, it achieves breakthroughs in coding and agent capabilities. Simultaneously, the open-source X2.6-4B on-device model was released, supporting three inference modes, offering improved inference efficiency compared to ollama, and available for free for a limited time. The X2.6 series is now open for trial on the official website.
The MiniMax Code desktop application has officially released the Mini App feature. This is a long-running plugin based on a Node.js backend and frontend pages, enabling bidirectional calls with Agents via hosted MCP, and supporting continuous operation in either the conversation sidebar or an independent window. This release includes the Mini App runtime, the /miniapp-creator rapid build tool, and official and community Mini App libraries; MiniMax Bot itself has been refactored into a Mini App and open-sourced to the MiniMax-Code-MiniApps repository.
A Huxiu author, using hands-on testing and cross-analysis of multi-source data, points out that the OpenRouter weekly leaderboard is a channel transaction list within a rolling 7-day window, not a global market indicator; topping the leaderboard is driven by pricing strategies (such as MiniMax's free period under an anonymous alias plus limited-time discounts), reflecting promotional intensity and subsidy generosity rather than model capability; the leaderboard has been used for stealth release pipelines, with eight anonymous aliases claimed seven times, essentially packaging the aggregator platform's own transaction data as industry credibility, and upgrading to a payment layer entry point via Stripe's acquisition.
Two undisclosed high-severity vulnerabilities have been disclosed in the underlying PHP language. The fix patch was originally scheduled for release on October 22 but was moved up to October 21 due to active exploitation by hackers. Rakuten stopped access to its seller dashboard starting at 22:00 and suspended front-end page display around 23:00, with the homepage showing a "System Maintenance" notice. Platforms such as Creema also entered maintenance mode, while Yahoo! Japan and Amazon Japan did not experience downtime. Cross-border sellers operating on Japanese sites need to monitor short-term disruptions in backend management and order fulfillment during the maintenance period.
Tesla announced that its Optimus humanoid robot has entered mass production at the Fremont factory in California, with the first batch of robots meeting mass production quality standards officially rolling off the line; however, Musk stated during the earnings call that external deliveries will be delayed until the end of 2027 due to AI capabilities, mechanical reliability, and key components such as hands not yet meeting commercial requirements, and there is currently no disclosure of orders or revenue.
A developer exploited the private key of the EM OV TLS CA – G3C intermediate certificate leaked by eMudhra (the original text states the private key is merely the number 4), redirected the cloud connection of the Yandex Station Mini 2 via DNS hijacking, forged an HTTPS certificate to deliver configuration enabling network ADB, and ultimately obtained Root privileges; this attack relies on the device having pre-installed and trusting the problematic root certificate, meaning older unupdated smart speakers, TVs, cameras, and other IoT devices may be affected.
SnapNeo is a local-first screenshot and screen recording tool for Windows, now available on the Microsoft Store. It supports visual region selection, multi-screen screenshots with annotations, scrolling long screenshots, area/window screen recording (MP4/GIF/WebP), local OCR (offline recognition with selectable and copyable text), desktop stickers (image/text/color cards supporting mouse click-through and color value conversion), smart masking (phone numbers/emails/faces/QR codes), real-time annotation and picture-in-picture during recording (Pro version), and historical screenshot retrieval and editing. Basic features are permanently free; the Pro version is a one-time purchase for permanent use. The official website is snapneo.app. Current system requirements are Windows 10 version 2004 or later and Windows 11 x64.
Qovia has released social media and e-commerce data interfaces for developers and AI agents, supporting over 20 platforms including Douyin, Xiaohongshu, Bilibili, Weibo, Zhihu, TikTok, and YouTube, with over 100 interfaces. It offers a unified JSON field structure, cross-platform search (POST /v1/search), and MCP integration capabilities, already adapted for Claude Code and Cursor, billed per successful response.
heybenben developed the RigCheck tool site (https://localrigcheck.com) to determine whether a GPU or Mac memory can run new open-source models locally.
NVIDIA expands the Cosmos platform with new training tools, simulation checkpoints, and Cosmos Data Factory blueprints; Cosmos Data Factory offers five cloud-native blueprints (autonomous driving data generation, image attribute enhancement, event video generation, defect image generation, and video data augmentation), supporting GitHub, Microsoft Azure, Nebius, and CoreWeave; Porsche, ANYbotics, and Ultralytics have already applied this toolchain to improve physical AI development efficiency; the Cosmos Framework supports fine-tuning and distillation of Hugging Face models such as GR00T, Alpamayo, QWEN, and Pi0.5; the Cosmos 3 Nano model variant supports multi-view, multi-sensor simulation strategy testing; the NVIDIA Metropolis VSS 3.3 blueprint implements video search and summarization based on Cosmos NIM microservices.
NVIDIA's official blog published a series of application cases for its technology in the European retail industry: Shopify uses NVIDIA Nemotron 3.5 Lightning, Megatron Bridge, and NeMo RL to reduce inference costs for Sidekick assistant sub-agents by 90% and improve response speed by 7 times; Zooplus achieves 86% search accuracy based on Nemotron Retriever and retail blueprints; Photoroom collaborates with Vestiaire Collective to process approximately 40,000 product images per day, resulting in a 5% increase in product liquidity and a 3%–7% increase in page views after using image editing APIs; Omi's ProductDrop workflow combines digital twin-generated images, a solution adopted by L’Oréal in the CreAItech project and integrated with ALCHEMI to accelerate ingredient screening; Mirakl runs open-source models on its own and European clouds such as Scaleway to ensure technological sovereignty; Otto Group deploys an autonomous warehouse simulation system based on Omniverse at its Haldensleben warehouse in Germany; ALLSIDES builds high-precision SimReady 3D digital twin factories in minutes; CATCHES' RealFit virtual try-on solution demonstrated a 55% increase in conversion rate, an 80% increase in add-to-cart rate, and a 56% increase in sales in AMIRI tests; Nebius showcased three retail blueprints at GTC Berlin: employee productivity agents, inventory simulation optimization, and agent fine-tuning flywheels.
NVIDIA has released StormScope EU, a new member of the Earth-2 open model family. This is a nowcasting weather forecast model trained specifically for Europe, based on EUMETSAT satellite observations, predicting cloud conditions at 10-minute intervals across 16 Meteosat channels; the model runs via Earth2Studio, with results generating map tiles in real time through Earthmover; AXA, Vaisala Xweather, University of Manchester, TotalEnergies and Space42 have already begun regional adaptation applications.
NVIDIA officially announced that its open models (including the Nemotron-Personas series and NeMo User Sim) are being used by Germany, Poland, the UK, France, Greece, and the European Commission to build controllable, localized, and compliant public sector AI systems; this involves upgrades to the deepset Haystack Enterprise platform, Poland's eHealth intelligent imaging platform, the UK Nemotron-Personas-UK dataset, Germany's Sinfonie voice interaction platform and sovereign cloud solutions, France's Twake.ai sovereign collaborative workspace, Greece's KIEFER Greek-language agent factory based on DGX SuperPOD, and the Codefest innovation competition.
The paper proposes the Agent4Sci test platform, covering 20 ML benchmarks across 11 scientific domains to evaluate the ability of frontier agents to improve published state-of-the-art models; experiments show that Claude Opus achieves an average relative improvement of 10.4%, but the gains mainly stem from ML engineering tricks rather than research-level insights; after adding lightweight plugin skills, Claude Opus improves to 20.1%, and Codex rises from 4.4% to 8.1%, with interventions becoming more domain-relevant; code and datasets are open-sourced.
This paper questions the SOTA accuracy reported by recent LLM-based node classification methods for text-attributed graphs (TAGs). By running four low-cost baselines (such as Correct & Smooth + TF-IDF, MLP + e5-large-v2 embeddings) on the same data splits, it finds that the median gap between these baselines and the highest reported mean of each paper is only 2.3–2.8 percentage points; in 15/24 tasks, the baseline gap is ≤3 points, but in 12/22 verifiable tasks, LLM methods are still significantly higher. The authors also point out multiple evaluation risks, including test set leakage, label misuse, undisclosed baseline reuse, and text-label misalignment in the Cora loader, and find that reported results for the same method on the same split vary by up to 12.5 percentage points across papers.
Viyozc has open-sourced a CLI tool named agent-shit-cleaner, which scans workspace directories, aggregates disk usage for each project, focuses on identifying and cleaning up dependency installation files and build artifacts generated during Agent development, and supports interactive operations to prevent accidental deletion.
The author deployed a Home Assistant system with 117 devices (including 105 Xiaomi devices) in their new home in Chengdu. The automation logic was primarily written by Claude Code (Opus 5.5). The article details the multi-condition trigger logic for 'Welcome Home', workarounds for unreliable Haier washing machine status, configuration for automatic valve closure upon water leakage, implementation of IPv6 relay via a J4125 side router, Tailscale remote access, lessons learned from AI misoperations, and shares tested recommendations for Tailscale combined with Ruijie AC+AP.
VMPal is a Mac virtualization tool optimized specifically for Apple Silicon chips, supporting one-click deployment of macOS, Windows 11, Ubuntu, and Fedora. It offers features such as Metal/Direct3D GPU acceleration, dynamic disks, snapshot rollback, and Parallels VM import. Its core innovation lies in providing an independent sandbox environment for AI agents, integrating with Claude Code, Codex, Grok CLI, and Gemini CLI via the built-in MCP protocol without requiring high-risk permissions on the host system. It only supports macOS 26 and above.
Zilu is a personal and family asset management tool released by an individual developer. It supports investment transaction records for stocks, funds, futures, etc., and provides customizable balance sheets, income and expense statistics tables, position analysis, and daily change charts for securities assets. The built-in AI Agent supports conversational financial reviews, stress tests, and health assessments, and enables quick bookkeeping through AI batch recognition of WeChat, Alipay, bank statements, and securities account statements. Currently accessible via the zilu.me website and WeChat Official Account, with data stored in the cloud.