Hermit is an Apache-2.0 framework that bolts a trainable external layer of instructions, skills, memory and tools onto an existing AI agent, trained from task feedback without changing the agent's model or harness.
Mango AI released a video tool that swaps multiple faces in group clips, aimed at creators editing scenes with several people.
Claude for Legal now offers artifacts for contracts and citation checks, letting legal users draft contract artifacts and verify cited authorities inside Claude.
Genspark launched AI Video Workbench, a tool that coordinates more than 70 AI models so beginners can produce short dramas and other videos from prompts.
Why it matters: A full-duplex conversational speech model trained purely on speech audio, from a fintech not a voice lab.
Block's AI lab released Dessa, a full-duplex speech-to-speech model trained only on human conversation audio. It runs locally on a MacBook Pro M4 Max, but Block published no size, benchmarks or weights.
Why it matters: Suggests much smaller speech models can match larger ones on transcription quality.
Nvidia's Nemotron 3.5 Transcribe is a speech-to-text model that Tech Times reports matches rival transcription models while using about one-sixth the active parameters.
Why it matters: Low per-character pricing intensifies competition in voice-agent speech APIs.
AssemblyAI-TTS-1 is a text-to-speech API for voice agents priced at $15 per 1M characters, aimed at builders adding spoken output.
ElevenLabs introduced Cue, a feature that gives character-level timing for generated audio, helping developers align speech with subtitles, captions and animation.
Why it matters: A rare open diffusion language model showing competitive agentic coding results.
Clover-35B-A3B-Preview is an open 35B-parameter mixture-of-experts diffusion language model built for agents and coding. DiffuSpace reports 73.6 on SWE-bench Verified and 515 tokens/s on one H200.
ESET launched AI security features aimed at consumers and small office/home office users, part of a broader push by security vendors to add AI-powered protection.
Xobin released AI Interview Auditor, a tool that reviews and checks AI-led job interviews so hiring teams can verify how candidates were assessed and make AI-driven hiring more trustworthy.
Why it matters: Tencent's flagship model adding visual understanding signals multimodal catch-up among Chinese labs.
Tencent's Dowson Tong said the full Hy4 and Hy4 Flash models will arrive soon with visual understanding, improving on the August Hy4 preview. Tencent also launched WorkBuddy agent across its office apps.
Feedoptimise now runs its AI product-feed optimization on its own GPUs and adds a free monthly AI quota to every subscription plan, aimed at retailers managing shopping feeds.
VoxNexus is a voice AI platform that runs across devices.
Why it matters: Shows a top Chinese model can execute multi-stage attacks, raising safety concerns.
Researchers say Z.AI's GLM-5.3 carried out 3 of 13 multi-stage cyberattack scenarios in Irregular's simulated tests, exposing safety review blind spots.
An Irish AI app helps families organise schedules, chores and logistics, aiming to reduce the chaos of coordinating family life.
Why it matters: Pushes agent-initiated payments into mainstream chat assistants, with approval controls.
PayBox by MoonPay lets users of ChatGPT and Claude authorize crypto transactions from within those assistants, adding a way for AI agents to initiate and approve on-chain payments.
Anthropic's Claude now produces smart reports that summarize how employees use it, including flagging work deemed off-topic for managers. Mashable reports on the feature and its workplace implications.
Google's Gemini can now generate interactive 'immersive explorations' that turn complex topics into playful, encyclopedia-like learning experiences, per Android Authority.
China released its first Fengyun weather satellite AI models for extreme weather: one interprets typhoon cloud imagery every 15 minutes, the other forecasts severe convection hours ahead.
DWS is an open-source command-line tool letting people and AI agents work with DingTalk chats, meetings, docs, calendars and approvals. An OpenHarmony build validated on HarmonyOS PCs was added.
Apex-35B-V1.0 is a model built on Qwen3.6-35B-A3B with help from Apex Intelligence's Apex Research agents. Internal tests show a 32.1-point gain on CFD-Bench.
KT has turned AI tokens into a telecom-style metered utility with an enterprise platform spanning some 500 models, letting companies measure, route and pay for model usage by the token.
SK AX released G-Fit Studio, a generative AI solution that runs work-based simulations to assess applicants' problem-solving, communication and AI skills, reportedly evaluating 1,800 candidates and interviewing them in six hours.
BallotBench gives nine AI agents internet access and $500 each to forecast 863 races in the 2026 midterms, publishing their research traces and revisions.
Expedia Group's fall 2026 release adds AI-powered tools and services intended to simplify the whole trip, from planning and booking through payment and customer support.
Why it matters: Signals how a major Chinese platform is justifying AI capex and extending AI into cross-border payments.
Tencent's cloud boss says AI demand justifies heavy capital spending as the company works to close its gap with rivals. Separately, Tencent and Visa expanded their partnership to explore AI-powered cross-border payments.
AI and data specialist WISE iTech entered the YouTube analytics market, applying its business analytics expertise to YouTube data with a subscription service.
CIIL Anuvani is an open speech AI project supporting 22 languages, released in 2026.
runkora is an open-source control plane for AI workloads; ClearRoute released it to manage AI deployments.
MoneyChoice Capital is making its economic forecasting data available to AI assistants through a Model Context Protocol server, letting agents query the data directly.
Why it matters: Monetizing AI search answers is a central test for Google's ad business as AI Mode grows.
Google is scaling ads in AI Mode to 10 additional countries including Australia, Canada, India, Kenya and Nigeria, and launching multiturn ads in the US that appear in follow-up queries.
Articul8 launched Bivia, a product that gives enterprise AI agents domain-specific decision intelligence so they can answer and act accurately within specialized business domains.
Purple is an AI agent that turns prompts into cosmetic formulas, aimed at R&D teams in the beauty industry.
Exoproc is an open-source MCP server over stdio that connects Claude Code to a live process's memory, so the model can inspect runtime state instead of simulating execution from decompiled code.
Digitate's ignio Hummingbird now combines governed IT agents with MCP integration, letting an AI assistant move from explaining a system failure to carrying out an authorized fix and verifying it worked.
Why it matters: A key read on whether massive AI capex is translating into cloud revenue.
Microsoft reported better-than-expected cloud revenue, with shares up about 5%, though heavy datacenter spending dampened cash flows and growth only accelerated slightly.
Deep Hat Sift is a model that decides which incoming alerts or logs an AI agent should investigate, avoiding costly uneventful runs at enterprise volumes.
Insurify opened its insurance marketplace to AI agents through a controlled gateway, aiming to preserve consumer choice and carrier protections while letting agents shop for coverage.
Cloudera introduced preconfigured AI applications built with Nvidia technology and expanded partnerships with Neo4j and Mistral AI to help enterprises move AI into production.
Ready Server launched ReadyServer LLM API, giving every VPS customer private AI inference at no additional cost.
Why it matters: Prefab rack designs and Rubin spec changes shape how fast and reliably AI datacenters can be built.
Nvidia announced DSX POD prefabricated server racks aimed at faster AI factory construction. It also updated Vera Rubin specs, saying the changed numbers reflect what systems sustain in volume production.
A startup founded by former Oppo executives unveils AI glasses aimed at travelers, with a different take on the smart-glasses form factor.
Megatron-MUSA-Titan is Moore Threads' open-source training stack for trillion-parameter MoE pretraining on 10,000-GPU clusters, with blockwise FP8 and support for slime and miles RL.
Iluvatar CoreX gave its Tiangai 300 general-purpose GPU its first live demo and says third-party tests put single-card throughput at 1.05 times Nvidia's H200. The test method was not disclosed.
Max Lo, founder of GEOSAT Aerospace & Technology, launched AerosoarX to shift Taiwan's drone sector from standalone manufacturing toward international defense-system integration, targeting US and Middle East supply.
Why it matters: Full-color microLED display brings AI glasses closer to mainstream everyday use.
Rokid unveiled Rokid Color, full-color AI glasses with a microLED optical engine, alongside Rokid Glasses 2 and the Rokid Ring wearable, starting at CNY4,000 in China.
Why it matters: Record Samsung profit shows AI memory demand now drives the semiconductor cycle.
Samsung Electronics reported record quarterly operating profit, topping 100 trillion won for the first time, as AI memory demand surged. Its chip division's 15-fold profit jump still missed estimates, and the device division stayed in loss.
A US company launched modular, Lego-like building kits for constructing offices, security checkpoints and robotics or drone facilities.
Why it matters: Consolidation in AI data center optics signals the next battleground beyond compute.
Marvell Technology is reportedly in advanced talks to buy data center optics maker Lumentum, a potential megadeal as chipmakers race to scale up AI infrastructure.
Amazon announced smart home updates including an Alexa-powered interactive 3D map of a home's interior and exterior, among its biggest announcements of the day.
Maxinsights raised $30M in seed funding for a platform that pays people to record everyday activities like cooking and cleaning, generating training data for robots and AI models. Founders are robotics and autonomous-driving veterans.
Altera announced a new Agilex FPGA portfolio plus an agentic AI developer tool aimed at helping developers build and deploy AI designs on its FPGAs.
Amazon's new Alexa Photo Frame lets users ask the voice assistant to display favorite photos. ZDNET reviews why the device fills a gap for the reviewer.
Why it matters: 512GB of unified memory makes the Mac Studio a serious local-model workstation.
Apple's Mac Studio with M5 Ultra is now available with 512GB of unified memory starting at $22,299, a configuration local LLM users have awaited.
AIMOGA Robotics, an OMODA and JAECOO ecosystem partner, unveiled its first home companion robot, moving automotive perception and control tech into homes.
Why it matters: Adds another Android XR entrant to the AI glasses race against Meta and Apple.
XREAL's Aura smart glasses run Android XR with Google Gemini built in, bringing AI assistance to a wearable display. Pricing is high, the report says.
PsiQuantum developed a device that routes signals between GPUs using optical links, which it says could exponentially speed up AI data center processing, per Semafor.
Samsung is partnering with Lark Health to offer eligible US seniors AI diabetes prevention coaching through Samsung Health, with Medicare eligibility restrictions.
Pakistan's Peshawar Safe City project has linked drone and anti-drone technology into its central command for city-wide surveillance and security, the first such integration in the country.
Why it matters: TPU distribution to rivals and hyperscalers could loosen Nvidia's accelerator dominance.
Google is widening AI chip sales through partners, with Visko running a live AI video model on Cloud TPUs and Oracle tapping Google TPUs as hyperscalers hedge Nvidia exposure. Google won't say whether the chip business is profitable.
OrcaRouter's Orca Connect turns any local GPU into a free, globally accessible AI API, the company says.
About Amazon · 21h · 3 sources
#
Why it matters: Amazon's Alexa+ push into camera-equipped home devices expands ambient AI but raises privacy concerns.
Amazon refreshed its Echo lineup: Echo Dot gains a built-in smart home hub and Wi-Fi extender, Echo Studio adds a camera for Alexa+ visual help, and Echo Show 16 offers a larger display. Privacy questions remain about the Studio camera.
Why it matters: First big enforcement move against AI wearables in high-stakes exams, setting a policy precedent.
South Korea's education ministry will seek criminal punishment and a three-year exam ban for CSAT test-takers caught using AI glasses, as proctors brace for hard-to-spot smart eyewear in exam halls.
Why it matters: First public financials from a leading humanoid maker gauge real demand for robots.
Unitree Robotics posted quarterly revenue growth of about 70% and 56% growth across the first nine months of 2026 in its first report since going public, a test for China's robotics sector.
An open-source Rust file launcher that searches inside documents, code and screenshots locally, using OCR, semantic search and MCP agent access. It drew 101 stars in its first week.
Gemini in Google Sheets is getting faster and adding source evidence for its answers, aimed at making spreadsheet AI output easier to verify and trust.
Hermit is an Apache-2.0 framework that bolts a trainable external layer of instructions, skills, memory and tools onto an existing AI agent, trained from task feedback without changing the agent's model or harness.
OpenAI now lets users message their ChatGPT agent by text to have it carry out tasks, extending agent control beyond the app interface.
A GitHub agent skill that examines a product you are authorised to rebuild, builds a clone, then tests it against the original and fixes gaps until a checklist passes. MIT licensed, about 100 stars its first week.
Ollama, a tool for running large language models locally on your own computer, shipped version 0.40.13 for macOS, Windows and Linux. The update lets users download and run open models offline without cloud services.
Amp released new versions of its command-line coding agent, installed with npm. Amp is a frontier coding agent and development environment that runs in the terminal to help developers write and edit code.
An Irish AI app helps families organise schedules, chores and logistics, aiming to reduce the chaos of coordinating family life.
Google's Gemini can now generate interactive 'immersive explorations' that turn complex topics into playful, encyclopedia-like learning experiences, per Android Authority.
DWS is an open-source command-line tool letting people and AI agents work with DingTalk chats, meetings, docs, calendars and approvals. An OpenHarmony build validated on HarmonyOS PCs was added.
A GGUF build of LiquidAI's 1.7B small language model with safety alignment ablated via a baked-in rank-1 LoRA, for running locally and offline in the Atomic Chat app.
A Visual Studio update adds Copilot chat modes and bulk merge conflict resolution, NewsBytes reports.
Why it matters: A whole-product benchmark shows coding agents still fail most end-to-end builds.
Replit released ViBench 2.0, a benchmark testing AI agents building 17 products from scratch and checking finished apps in a browser; top models pass 49%-65% of whole products.
37studio is a self-hosted open-source alternative to Buffer and Opus Clip: Claude Code makes on-brand videos, posts them to nine platforms and drafts replies, using your own API keys.
runkora is an open-source control plane for AI workloads; ClearRoute released it to manage AI deployments.
Meta's developer MCP server gains six tools that let AI agents create and manipulate apps, extending agent capabilities on the Meta developer platform.
Articul8 launched Bivia, a product that gives enterprise AI agents domain-specific decision intelligence so they can answer and act accurately within specialized business domains.
Exoproc is an open-source MCP server over stdio that connects Claude Code to a live process's memory, so the model can inspect runtime state instead of simulating execution from decompiled code.
A Reddit thread suspects Anthropic uses Claude to iteratively redesign the Claude Desktop app, breaking and restoring features between releases.
Digitate's ignio Hummingbird now combines governed IT agents with MCP integration, letting an AI assistant move from explaining a system failure to carrying out an authorized fix and verifying it worked.
LiteRT in Google Play services now offers GPU inference via ML Drift, about 1.4x faster, with a new CompiledModel API. Switching from bundled binaries cuts app size by roughly 8 MB and enables automatic runtime updates.
A Cohere tutorial on AWS shows how to build a financial research assistant using Cohere's models and AWS services, likely combining retrieval and generation for analysts.
NamiGPT is a beginner-friendly GitHub tutorial that walks through writing GPT-2 from scratch in PyTorch line by line to show how large language models work.
An AWS blog walks through architecture and an open-source reference blueprint for a governed, multi-account agentic AI platform on Amazon Bedrock AgentCore, covering control plane, gateways and CI/CD evaluation.
Why it matters: Cutting token use per coding task targets the biggest cost driver for agentic software work.
Particle-1 is a post-trained version of DeepSeek V4.1 Flash, offered via API, that uses fewer tokens across four software and data-engineering benchmarks, cutting cost per coding task by 34%.
An AWS blog walks through fine-tuning Qwen2.5-1.5B with PyTorch FSDP across eight GPUs on two EFA-enabled g6.12xlarge instances on Amazon EKS using the Ray Train Deep Learning Container.
Part 3 of an AWS series on QA Studio with Amazon Nova Act adds apps that organize tests and report health metrics, plus team-based user management controlling test ownership.
Google's AI Mode in the US declines to say who won the 2020 election, while Gemini and versions abroad will answer; the discrepancy drew scrutiny on Hacker News.
A Google Developers post walks through the Agents CLI evaluation loop: continuous testing and grading to catch prompt regressions, compare models via side-by-side trace comparisons, and automate evaluation passes.
Reindeer joined Microsoft's Pegasus program and listed its managed, single-tenant platform that connects agents to systems like SAP and Oracle on Azure Marketplace, with eligible purchases counting toward Azure commitments.
A GGUF build of LiquidAI's 1.7B small language model with safety alignment ablated via a baked-in rank-1 LoRA, for running locally and offline in the Atomic Chat app.
Fibo 2.0 Edit Turbo is an instruction-based image editing model that edits images in four guidance-free steps, using up to four reference images and an optional mask; a Hugging Face Space demo is available.
This Hugging Face Space runs Wan 2.2 image-to-video generation in a browser demo, turning a still image into a short clip. It has about 100 likes and is a community-built Gradio app.
A merged image/video model built by blending two Minimax-H3 Ref2VA finetunes with a custom PRISM algorithm, no mergekit. Weights shared in fp16 on Hugging Face.
Clover-35B-A3B-Preview is DiffuSpace's first model: a 35B-parameter mixture-of-experts diffusion language model with about 3B active parameters, claimed state of the art among open diffusion LMs.
Tencent Hunyuan's Hy-SimulMT-30B-A3B-FP8 is a low-latency, high-quality model for real-time simultaneous translation, released under Apache 2.0 on Hugging Face alongside the Hy-Nomas model.
MiniCPM-OCR is an open document-reading model that also checks whether its own OCR output is trustworthy. OpenBMB released the 1.3B MiniCPM-OCR 1.0, scoring 93.68 on Real5-OmniDocBench for curved and real-world pages, plus an AI judge model.
Liquid AI released LFM3, a family of six open-weight 2B-class multimodal models for building specialized models, available on Hugging Face.
VDEBench is a Hugging Face dataset of 3,506 examples testing instruction-guided image editing on dense Chinese and English document pages, where models must change requested text or tables while preserving layout and typography.
An open-source Rust file launcher that searches inside documents, code and screenshots locally, using OCR, semantic search and MCP agent access. It drew 101 stars in its first week.
A GitHub agent skill that examines a product you are authorised to rebuild, builds a clone, then tests it against the original and fixes gaps until a checklist passes. MIT licensed, about 100 stars its first week.
37studio is a self-hosted open-source alternative to Buffer and Opus Clip: Claude Code makes on-brand videos, posts them to nine platforms and drafts replies, using your own API keys.
NamiGPT is a beginner-friendly GitHub tutorial that walks through writing GPT-2 from scratch in PyTorch line by line to show how large language models work.
AgentFlux is a platform for building visual workflows and autonomous AI agents, runnable with Docker on your own computer or server. It drew 101 GitHub stars in its first week.
Kanzo is a self-hosted kanban board for both people and AI agents, shipped as a single 4 MB Rust binary that starts in 2 ms with zero setup.
TokenBar is an open-source, cross-platform app that monitors AI usage quotas for Codex, Claude, Cursor and 47 other services. Built with Flutter for Windows, macOS, Linux, Android and iOS.
goose is an open-source, extensible AI agent that installs, executes, edits and tests code with any LLM. Version 1.57.0 opens the release branch for the next update.
BLA (Blind LLM Agents) is an open-source adversarial defense that protects binaries, Electron apps, JS bundles, Python and WebAssembly against AI reverse-engineering agents by poisoning, exhausting and hallucinating them.
anymove is an open-source Python project that captures motion for any creature from one AI-generated video: it reads joints with DINOv2 across two cameras in one frame and outputs a fitted GLB rig.
The Deep View profiles Halo AI glasses, which emphasize radical openness and privacy; adjacent items cover Claude for Mac 4.29.0, ChatGPT for Mac updates and Gemini CLI 0.66.0.
Gemini in Google Sheets is getting faster and adding source evidence for its answers, aimed at making spreadsheet AI output easier to verify and trust.
OpenAI now lets users message their ChatGPT agent by text to have it carry out tasks, extending agent control beyond the app interface.
An Irish AI app helps families organise schedules, chores and logistics, aiming to reduce the chaos of coordinating family life.
Google's Gemini can now generate interactive 'immersive explorations' that turn complex topics into playful, encyclopedia-like learning experiences, per Android Authority.
A Reddit thread suspects Anthropic uses Claude to iteratively redesign the Claude Desktop app, breaking and restoring features between releases.
Google's AI Mode in the US declines to say who won the 2020 election, while Gemini and versions abroad will answer; the discrepancy drew scrutiny on Hacker News.
Unstoppable Domains launched Unstoppable.ai, a model-agnostic AI workspace that lets users work with multiple AI models in one place.
At its Berlin DevDay, OpenAI said its always-on Dots agent will roll out more widely in Europe in a few weeks, and is expanding ChatGPT dot access to iMessage and RCS. Weekly Codex users in Germany grew twentyfold since January.
Raycast, a keyboard-driven launcher for Mac, Windows and Linux that lets you control tools with a few keystrokes, shipped versions 2.10.0.0 and 2.10.1.0.
Monologue 1.5.3 is AI voice dictation for macOS, Windows and Linux that adapts to your writing style; Superwhisper also shipped version 3.1.0 of its dictation app.
Nvidia released the RTX Remix 1.6 update, which adds path tracing support in Minecraft.
Google's AI Overviews in Search cannot be disabled, but PCMag offers workarounds for users who want to avoid the AI-generated summaries.
Otter added a Work mode to its AI Chat assistant that automates more tasks before and after meetings, part of its push beyond transcription following a $320 million funding round last week.
Klaviyo released an AI agent that generates email marketing automation flows from a prompt, aimed at its marketing customers. It turns plain-language instructions into ready-to-run flows.
Digit reports Gemini could soon handle eSIM setup for users, simplifying the process of activating mobile plans on a phone.
Eney is an AI assistant that understands what is happening across a user's apps, files and system, prepares next steps and asks before acting. It offers copy-triggered suggestions and bot-free meeting transcription.
Clearafi compares answers from multiple AI models and scores their agreement before you rely on the result; the tool costs $50.
Zocdoc now lets registered AI agents search providers, check insurance and book appointments, beginning with a partnership with Instinct.
Chrome's enhanced autofill now uses Gemini and Gmail to fill forms, rolling out as 'Suggested by Gemini.' Google also added more Connected Apps to Gemini and Search and opened a new developer platform for third-party app integration.
The Washington Post put chatbots to the test to see whether they can reliably create voter guides, assessing trustworthiness and accuracy.
Rori, an AI math tutor, has reached roughly 250,000 students across Liberia, Rwanda, Ghana and Sierra Leone, with parents embracing the tool.
Adalat AI says its paperwork tools are used in 6,000 courtrooms across 11 Indian states, speeding up the justice system.
Waddle World is a Tamagotchi-style pet app whose virtual creatures chat with users and doodle illustrations for conversations. Meta describes it as a daily journal disguised as a pet; it is in testing in the Philippines.
WIRED reports that AI chatbots could help with emergency communication, since humans struggle with it, but only if people take the AI alerts seriously.
Users reported Microsoft Copilot issues including errors with AI generation, the app and website, and repeated 'Try Again' prompts, according to outage trackers compiled by The Sunday Guardian.
Hermit is an Apache-2.0 framework that bolts a trainable external layer of instructions, skills, memory and tools onto an existing AI agent, trained from task feedback without changing the agent's model or harness.
Ollama, a tool for running large language models locally on your own computer, shipped version 0.40.13 for macOS, Windows and Linux. The update lets users download and run open models offline without cloud services.
Amp released new versions of its command-line coding agent, installed with npm. Amp is a frontier coding agent and development environment that runs in the terminal to help developers write and edit code.
DWS is an open-source command-line tool letting people and AI agents work with DingTalk chats, meetings, docs, calendars and approvals. An OpenHarmony build validated on HarmonyOS PCs was added.
A Visual Studio update adds Copilot chat modes and bulk merge conflict resolution, NewsBytes reports.
Replit released ViBench 2.0, a benchmark testing AI agents building 17 products from scratch and checking finished apps in a browser; top models pass 49%-65% of whole products.
runkora is an open-source control plane for AI workloads; ClearRoute released it to manage AI deployments.
Meta's developer MCP server gains six tools that let AI agents create and manipulate apps, extending agent capabilities on the Meta developer platform.
Articul8 launched Bivia, a product that gives enterprise AI agents domain-specific decision intelligence so they can answer and act accurately within specialized business domains.
Exoproc is an open-source MCP server over stdio that connects Claude Code to a live process's memory, so the model can inspect runtime state instead of simulating execution from decompiled code.
Digitate's ignio Hummingbird now combines governed IT agents with MCP integration, letting an AI assistant move from explaining a system failure to carrying out an authorized fix and verifying it worked.
LiteRT in Google Play services now offers GPU inference via ML Drift, about 1.4x faster, with a new CompiledModel API. Switching from bundled binaries cuts app size by roughly 8 MB and enables automatic runtime updates.
A Cohere tutorial on AWS shows how to build a financial research assistant using Cohere's models and AWS services, likely combining retrieval and generation for analysts.
AWS Machine Learning · 14h
#
An AWS blog walks through architecture and an open-source reference blueprint for a governed, multi-account agentic AI platform on Amazon Bedrock AgentCore, covering control plane, gateways and CI/CD evaluation.
Particle-1 is a post-trained version of DeepSeek V4.1 Flash, offered via API, that uses fewer tokens across four software and data-engineering benchmarks, cutting cost per coding task by 34%.
AWS Machine Learning · 14h
#
An AWS blog walks through fine-tuning Qwen2.5-1.5B with PyTorch FSDP across eight GPUs on two EFA-enabled g6.12xlarge instances on Amazon EKS using the Ray Train Deep Learning Container.
AWS Machine Learning · 14h
#
Part 3 of an AWS series on QA Studio with Amazon Nova Act adds apps that organize tests and report health metrics, plus team-based user management controlling test ownership.
A Google Developers post walks through the Agents CLI evaluation loop: continuous testing and grading to catch prompt regressions, compare models via side-by-side trace comparisons, and automate evaluation passes.
Reindeer joined Microsoft's Pegasus program and listed its managed, single-tenant platform that connects agents to systems like SAP and Oracle on Azure Marketplace, with eligible purchases counting toward Azure commitments.
MiniMax released MCode Stack, a set of 15 open agent skills its own coding team uses daily, for developers building AI coding agents.
TaxDome released a public API and an MCP server letting AI agents and developers connect to its tax and accounting practice management platform.
Google's Bring Your Own Agent feature lets Android Studio run over 40 coding agents, including Anthropic's Claude Agent and OpenAI's Codex, plus terminal tools Google says finish UI tasks 77% faster.
Controlled GitHub tests found Claude Code client labels use five-hour quotas differently, including a roughly 2.9x gap in one Team-plan test, and a three-word status line can burn 15-20% of tokens.
GitHub touted Copilot Fusion for frontier-level AI coding at half the cost, made Agentic Workflows generally available, added new triggers and cost controls for Copilot Automations, and previewed code scanning for AI applications.
OpenMausBot, published on GitHub by Milind Soni, is an open-source personal agent chat app where each bot is a local Claude or Codex agent with its own personality, model, cloud computer and connected apps, used like messaging contacts.