Google and the Meta PyTorch team have jointly open-sourced TorchTPU, a native TPU software stack that allows developers to deploy PyTorch models to Google TPUs with minimal code changes; it has been validated for running Qwen3-8B training on up to 8000 Ironwood TPUs, achieving a maximum model FLOPs utilization of 52%; Indian AI company Zenteiq.ai saw inference throughput increase by more than 4x after using the solution; TorchTPU is integrated with tools such as vLLM, SGLang, llm-d, TorchTitan, VeRL, Ray, Hugging Face Transformers, and LiteRT, supporting training, inference, reinforcement learning, cluster scheduling, and edge deployment; the open-source address is https://github.com/google-pytorch/torch_tpu.
The AMD ROCm official blog released Micro-World 2, a causal autoregressive world model based on Wan2.2 5B, supporting minute-level real-time video stream generation controlled by keyboard/mouse, achieving 32.12–42.49 FPS on a single AMD Instinct™ MI350X GPU; it was trained using 440 frame-aligned action trajectories collected via the Capsaicin framework, with all weights, training code, inference code, and datasets open-sourced, and deeply optimized for the ROCm ecosystem.
Personal Agent products such as Meta Muse, xAI Grok, OpenAI Dots, Instinct, and ByteDance's Doubao mobile assistant have universally encountered refusals of automated operations from 155 apps including Amazon, Yelp, Delta Air Lines, JPMorgan Chase, and domestic apps like WeChat, Taobao, Alipay, Kimi, and DeepSeek; the core contradiction lies in platforms' concerns over lost ad exposure, transaction commissions, and data sovereignty, with security serving only as a superficial justification; Meta has shifted to targeted partnerships with retailers like Walmart, while Alibaba chooses to close-loop task handling within its own ecosystem. The industry is promoting personal Agent protocols and national standards, but general-purpose cross-platform Agents are difficult to implement in the short term.
ManageEngine Zoho has released over 30 Zia intelligent agents for the Chinese market, powered by its proprietary Zia large model. These agents generate no additional token costs and are directly integrated into customers' deployed IT management products, supporting full-stack IT operations scenarios such as root cause analysis, application monitoring, and service desk ticket processing. The agents operate within local data centers in Beijing and Shanghai, ensuring data remains within the domain and is not used for model training. The company plans to expand to over 100 agents by the end of 2026, with an official debut scheduled at the Shenzhen User Conference on October 28.
AMD has launched an agentic AI assistant named Ross, targeting the full embedded development workflow. It supports natural language interaction and integrates four major components: MCP servers, the AMD knowledge base, expert-authored agent skills, and design examples. It is currently accessible via AMD.com/Ross, with support for Vivado and Vitis HLS tools; power design, FPGA debugging, and AI inference capabilities will be added in Q4. Examples show that it can reduce a Vitis HLS image sensor algorithm optimization task from 2 days to 2 hours and decrease programmable logic resource usage by 35%.
DeepSeek released the V4.1-Flash-1017 minor version update, featuring a 552 billion parameter MoE architecture focused on programming Agent scenarios; single-task costs dropped from 3.3 yuan to 1.3 yuan, and Coding Agent capability (DeepSWE) improved from 74.2% to 77.0%, but general reasoning (HLE) showed no significant improvement; The Information reported that DS plans to use Ascend 950DT for frontier model pre-training by the end of 2026, facing engineering challenges such as communication attenuation in ten-thousand-card clusters, stability, and MFU; the US DHS ban prohibits federal agencies from procuring third-party products based on DeepSeek, while China intends to issue an AI technology export catalog, imposing controls on overseas deployment of models with over 2 trillion parameters or top-ten global capabilities, primarily squeezing the 'third-party cloud hosting + licensing revenue sharing' pathway.
Eon replaced scattered per-tool credentials and local MCP processes with the Amazon Bedrock AgentCore Gateway to build a centralized, Cedar-policy-driven MCP gateway. This enables developers and autonomous SRE agents (running on Claude) to have unified, least-privilege access to multi-cloud investigation tools across AWS/GCP/Azure (such as Grafana, GitHub, and Kubernetes APIs). Onboarding time for new engineers was reduced from two days to two minutes, local MCP processes were eliminated, and cross-account connections rely solely on IAM role trust policies.
Agent Archive is an open-source CLI tool that released version v0.2.2, supporting session context handoff between Claude Code, Codex, and Cursor, and automatically uploading conversation records to the user's own S3 or R2 storage buckets; it provides commands such as list, search, show, handoff, and stats, enabling synchronization of session history across multiple machines, supporting macOS and Linux (x86-64/arm64) platforms, licensed under MIT, with source code hosted on GitHub.
This article is a hands-on tutorial published by the AWS Machine Learning Blog, demonstrating how to use the Denodo Platform (deployed via Agora or EC2/EKS) to build a customer 360 semantic layer spanning Amazon Aurora, Amazon Redshift, and on-premises PostgreSQL. It exposes this layer as MCP tools through the Denodo MCP Server and connects it to Strands agents hosted by Amazon Bedrock AgentCore. The tutorial covers base/derived/metric view modeling, dynamic row/column filtering via Global Security Policy, On-Behalf-Of identity pass-through, details of VQL query generation and federated execution, and results from three types of validation queries (single-source, dual-source, and triple-source).
Individual developer weishao666 released the image-to-editable-PPTX service image2ppt.asia. It uses an Agent loop to achieve element-by-element reconstruction: text becomes native text boxes, shapes become native shapes, and photos become independent layers. Internally, it used 62 real EdrawMax images for regression testing, delivering 60/62 successfully; text recall ≥0.95 was achieved in 61/63 runs, and pixel difference ≤5 was achieved in 62/63 runs. Processing a single image takes 20–40 minutes. Currently, very large images (>15MP) and long strip images are unstable.
This article is the third tutorial in the AWS Machine Learning blog series, demonstrating how to use Strategy Mosaic as a semantic layer, combined with Strands agents hosted by Amazon Bedrock AgentCore, to implement Customer 360 queries across Amazon Redshift and Amazon Aurora.
This article is the third in the AWS Machine Learning Blog tutorial series, demonstrating how to use AtScale (version C2026.7) to build an agentic semantic layer on Amazon Redshift Serverless and integrate it with Amazon Bedrock AgentCore via the Model Context Protocol (MCP).
The Mac version of Perplexity Computer has added an on-device mode, supporting two models: Qwen 3.6 and PPLX Pro. It can perform task planning, reasoning, and file processing on the device, accessing local files or uploading to the cloud only after user authorization. This feature is available to Pro, Max, and Enterprise subscribers, with a minimum requirement of an Apple chip and 36GB unified memory.
IBM Research and Hugging Face jointly released the ALTK-Evolve method, which uses GPT-4.1 to automatically extract short behavioral rules (such as user authentication and refund confirmation processes) from its own training trajectories. These rules are appended to the ReAct agent system prompt in a pure prompt-based manner, without updating model weights or requiring human annotation. On the retail and airline domains of the τ²-bench customer support benchmark, this approach achieved Pass^4 scores of 67.5% and 45.0% respectively for GPT-4.1; the retail score surpassed the fine-tuned model RAFT-30B-A3B and all frontier company models. The GPT-5.5 version achieved a Pass^4 score of 82.0% in the airline domain, ranking first on the leaderboard. All experiments were based on the official implementation of τ²-bench, using the same model for both guideline extraction and execution.
Nimbalyst Wiki is a free plugin that supports any MCP agent, including Claude Code, Codex, Cowork, Cursor, and VS Code. It automatically generates Markdown files, diagrams, CSVs, and clickable HTML pages written by the Agent within local repositories; users can edit directly in the browser without installing an application or registering an account.
Penmentor is an xournal++ plugin that supports real-time annotation of handwritten mathematical content by invoking Claude, Gemini, or GPT via keyboard shortcuts. It provides feedback with color markers, offers step-by-step hints instead of answers, and supports generating 3Blue1Brown-style manim interactive animations from selected formulas. It is based on users' existing AI command-line agents (such as Claude Code, Codex, Antigravity CLI) and requires running in a Linux X11 environment with rofi, slop, xdotool, and PyQt6. The project is open-source under the MIT license.
PDFkind is a PDF editor that runs in the browser, supporting double-click editing of existing text while maintaining the original font, size, and color, with automatic line reflow; redaction permanently deletes content rather than covering it; provides tools including 20+ fonts, images, signatures, form fields (text/checkbox/radio/dropdown/signature), watermarks, headers, page numbers, etc.; all processing is done locally, files are not uploaded to servers, and edits are automatically saved on the device.
Twig is an open-source Android file manager. Version 1.14.0 has been released, supporting unified mounting of various storage sources such as NAS, S3, restic, and Git into a single tree with in-place expansion; implemented purely in Kotlin without Material library dependencies, including self-developed implementations for protocols like WebDAV/S3/restic/Git; features include dual-pane browsing, directory comparison, Wi-Fi sharing, built-in terminal, Jellyfin/Emby media library integration, and customizable website bookmarks; APKs are available in full/libre/single-architecture versions with consistent signing keys, supporting automatic updates via F-Droid and Obtainium; the project uses the GPL-3.0 license, with code hosted on GitHub.
VoidAccess is a self-hosted open-source dark web open-source intelligence (OSINT) tool that supports converting dark web research queries into structured threat intelligence. It provides parallel collection, entity extraction, multi-source enrichment, relationship graph construction, and export in 12 formats including STIX, MISP, and Sigma.
The open-source project Clover Mail (clover-mail) has been released. It supports visual management of multiple domains and mailboxes via a Web panel on Debian/Ubuntu servers, integrates automatic Cloudflare DNS configuration for MX/SPF/DKIM/DMARC, includes Roundcube Webmail, and is compatible with mainstream email clients. It offers one-click installation scripts and Docker Compose installation methods, licensed under AGPL-3.0-or-later.
ElevenLabs has launched its Realtime series speech-to-speech models, including v1 Mini (for natural conversations) and v1 Max (for complex tasks), with first-audio latencies of 620ms and 570ms respectively; it supports 90+ languages and can be accessed via ElevenAgents or an OpenAI-compatible Realtime API.
Lykr has released the local AI image generation app Pictide (Hua Liu), which supports prompt-free operation, local preference learning based on user clicks of like/dislike, manual fixing of image elements, automatic combination of Danbooru tags, and integration with ComfyUI; the project is open source.
Maap is a 4B parameter model based on Qwen3-4B-Base, equipped with LoRA adapters and distribution heads, specifically designed to output calibrated intervals (such as 80% confidence intervals) and probabilities (such as prob_over). It covers 104 estimation targets from 36 public datasets. It offers the first 1,000 calls for free, with response times of approximately 160 milliseconds. The API returns fields such as median, quantiles, intervals, and novelty_flag, and explicitly labels the 44 targets where calibration failed.
Beatsflow is an AI tool for music producers that supports generating complete beats via natural language prompts (e.g., 'Midnight Drift Dark Trap & Drill 140 BPM C Minor') or by uploading MP3/WAV reference audio. It outputs 4-track stems (Drums/Bass/Melodic Chords/FX), multi-section dynamic arrangements (Intro/Verse/Chorus/Bridge/Outro), precise control over BPM and key (60–180 BPM, supporting major/minor keys), anti-clipping and low-frequency mud DSP processing, and generation in as little as 15 seconds.