Why it matters: A new lab model aimed at automating AI training work, plus a huge reported raise, signals a fresh frontier competitor.
Thinking Machines Labs released Inkling 2 Small, a large language model that automates AI training tasks. Business Insider separately reported the startup is raising about $5 billion.
NousResearch post-trained Alibaba's Qwen3.8-27B to reinforce Western values, releasing the resulting Arete-27B model on Hugging Face as an open-weights release.
Why it matters: Shows a 22 MB model can handle multi-task vision on-device, widening local physical AI.
Mirage is a tiny vision-language model that runs scene description, object detection, OCR, segmentation and spatial reasoning in a single pass. Its weights file is 22 MB, aimed at phones, wearables and robots.
Xing4.0-29B-A4B is a 29B mixture-of-experts model in the Xing series (formerly TeleChat) from China Telecom's AI division. bartowski published GGUF quantizations for local use.
AegisKernel, from a 16-year-old solo builder called Kshatriyag, sells deterministic tool checks for LangChain, CrewAI and AutoGen agents so hijacked agents are blocked before they run tools. It is aimed at agent security.
Higgsfield's Micro Drama Ads tool turns a product into a plot twist, adding eight styles and 44 recreatable story cards to Ads Studio. The same catalog is available through Claude and ChatGPT.
OpenEvidence will embed English and Spanish American College of Rheumatology patient education sheets into its medical AI platform for clinicians treating rheumatic conditions.
Saber is a new inference engine claiming 6,645 tok/s prefill and 316 tok/s decode peak for Qwen3.8 27B. Its developer says it takes a different approach from existing engines after a month of testing.
BMO introduced an agentic AI assistant that helps credit card customers get answers, accelerating support case preparation and refund processing.
Why it matters: A machine-checked proof of a known conjecture would mark a notable AI math milestone.
Physical Superintelligence says a system using a non-public Anthropic model produced a Lean-accepted, machine-checked proof of the Aaronson-Ambainis conjecture.
Kabilio raised €10 million to launch an AI agent that handles accounting work, aiming to automate bookkeeping tasks for businesses.
Rune is an open-source Rust UI framework that renders interfaces on the GPU and targets macOS, Windows, Linux and the web from one codebase.
Why it matters: Sub-200ms duplex voice-plus-video models push real-time avatars toward natural conversation.
Boson AI released Higgs Realtime Duplex, a 45B model that processes voice and video in 160ms chunks to converse with a rendered face, handle tools and delegate harder questions.
Why it matters: Gives the industry a shared yardstick for whether AI shopping agents actually work.
Visa released a benchmark evaluating how well AI shopping agents complete purchases, reporting test results and tying it to its agentic commerce standards push.
10x Genomics released Sentira, an AI tool meant to simplify analysis of single-cell data, lowering the expertise needed to interpret experiments.
Mayo Clinic Platform released an AI tool aimed at streamlining clinical decision-making for care teams.
A small, local text-to-speech model that runs without an internet connection in about 32 MB, shown off in a browser mini-game. It lets devices generate speech on their own hardware.
Dell Technologies is adding identity recovery capabilities to its AI security push, aiming to help businesses restore compromised accounts and credentials.
An OpenAI engineer charged with planting a hidden recording device in a company office pleaded not guilty in federal court.
Oc-e-Non-v1 is a fine-tuned Qwen3-4B-Instruct-2507 model that halves mean response length, aiming for more concise answers.
Safegres is a read-only verifier that grades Postgres grants and row-level security in CI, so teams can check database permissions before AI-built apps ship.
Vodia's phone system (PBX) now supports xAI's Grok assistant, letting businesses bring the chatbot into their phone workflows, Telecom Reseller reports.
Why it matters: Pushes AI coding agents from autocomplete into full app building and deployment.
Cursor's new Build Live Apps feature creates and deploys web applications with databases, authentication and hosting from prompts, TechRepublic reports, detailing what is included and what it costs.
Subaio released an MCP server that connects AI assistants to real-time bank account data, letting agents query a user's finances.
ICON launched FIRECREST, an AI video engine that generates video communications for clinical trial participants and communications.
Insilico Medicine released Longevity Grok, an open-source system that runs up to 30 agents on Grok Bot to organize fragmented health records, propose interventions and score later biomarker results.
Why it matters: A major bank adopting OpenAI for security and coding shows enterprise trust in AI for defense.
JPMorgan Chase signed a deal with OpenAI covering cybersecurity and coding, as SpecterOps joined OpenAI's partner network to bring Daybreak red-team capabilities to enterprise security teams.
Rex Africa launched an AI-native bank for Nigerian customers, centered on an agent called Rex Atlas. Details on products and availability were not specified.
Why it matters: Hybrid quantum-classical HPC could speed error-correction work, a key path to useful quantum computing.
IBM, AMD and Oak Ridge National Laboratory launched a hybrid quantum-classical HPC system to accelerate quantum error-correction research ahead of 2028.
Atomicwork's AI Coworker Marketplace offers prebuilt, certified AI coworkers for enterprise IT and security teams, letting agents work alongside employees.
Box Architect is a natural-language AI tool for building automated workflows around document processes. Announced at BoxWorks 2026, it acts as an agentic operator of Box's content and workflow capabilities.
Ro Research Lab combines AI, real-world patient data and scientific research; Ro plans to publish early findings, submit validated ones for peer review and turn evidence into clinical guidance and tools.
Why it matters: Largest private AI funding event sets a $1.25T valuation benchmark for the sector.
OpenAI closed roughly $11B of new funding, the first part of a round valuing it at $1.25T, with $9B more targeted. SoftBank is reportedly weighing an added investment under $1B.
Nivo launched a website letting lenders try its AI Assistant for Lenders free for two weeks with no commitment, aimed at supporting lending workflows through conversational AI.
JuvrixAI launched a no-code platform for building AI agents on NEAR Protocol, letting users assemble and deploy agents without writing code.
Wizstar released Wizshot 1.0, its first in-house AI video generation model, pitched as making high-quality AI video creation more cost-efficient.
Edera announced scope limits on credentials used by AI agents to reduce blast radius if an agent is compromised. Credential-free API access for agents is still in development.
Why it matters: An Anthropic listing near a trillion-dollar valuation would be a landmark test for AI company valuations.
Anthropic plans to push ahead with an IPO roadshow next week, with buyers weighing valuations from about $1.1T to $1.5T. Wall Street firm XN has amassed a $5.3B stake in the AI lab.
ElevenLabs' Ads Engine now includes human review from localization experts and creative specialists for on-brand localized ads, and ElevenMusic creations can be published directly to SoundCloud.
Alon Goren and David Bleznak launched LenDb, a free database of 224,480 loan officers built for AI assistants, with their internal AI agent Andrew running daily operations.
Boston Dynamics' new Stretch Gen 2 warehouse robot took two and a half years to develop and targets more reliable, consistent logistics performance, per its logistics VP.
Lockheed Martin's Lamprey Multi-Mission Autonomous Undersea Vehicle can recharge its batteries while attached to host ships, extending the time it can operate underwater.
Aina's $399 AI ring lets users approve tasks with a thumb flick, with iPhone and Mac compatibility and shipping plans aimed at both consumers and IT teams.
AWS Machine Learning · 5h · 2 sources
#
Why it matters: GPU partitioning lets cloud users pack more models per accelerator and cut inference cost.
Amazon SageMaker Inference now supports Nvidia Multi-Instance GPU, splitting one GPU into isolated partitions so users can run separate models or right-sized fractions of a newer accelerator per endpoint.
Lightweight industrial drone ships globally with standard LiDAR, millimeter-wave radar, omnidirectional vision, 176x zoom and 59-minute flight time.
Why it matters: Signals AI compute demand reassessment and pressure on neocloud providers in the buildout.
Microsoft is winding down an arrangement to rent AI computing from Oracle as it reassesses billions of dollars in cloud partnerships, and plans to rely on fewer neoclouds as the AI buildout shows strain.
Why it matters: Gives AMD GPU users faster, lower-error attention kernels as an open alternative to AITER.
Zyphra open-sourced FP8 and BF16 prefill attention kernels for AMD Instinct MI355X that run 1.16-1.76x faster than AMD's AITER, with 62-78% lower error at head dimension 128.
Why it matters: Per-token post-training billing lowers the barrier to fine-tuning open models.
Baseten's Serverless Loops, a managed SDK for supervised fine-tuning and reinforcement learning of open models with per-token training billing, is now generally available and connects checkpoints to Dedicated Inference.
Why it matters: Signals OpenAI's capex commitments survive political uncertainty in the US.
OpenAI CFO Sarah Friar told CNBC the company will keep investing in AI data centers regardless of midterm election results, pressing ahead with its buildout plans.
Why it matters: Humanoid robots doing paid real-world work with teleoperation is a step toward autonomous service robots.
Tau Robotics offers humanoid robot home cleanings at $60 a visit, down from earlier pricing, with a human remote driver steering the robots to gather training data.
After criticism of Copilot being forced on users, Microsoft says its addition to the Windows 11 search interface will be optional, Wccftech reports.
Why it matters: Memory is a growing AI bottleneck; TSMC entering it could reshape the supply chain.
TSMC is considering expanding into advanced memory manufacturing, including 3D DRAM, which would give the world's largest chipmaker a bigger share of the AI hardware supply chain, per The Information.
Why it matters: Targets governments and enterprises wanting agentic AI inside sovereign clouds.
Tencent Cloud Enterprise launches a full-stack intelligent private cloud aimed at sovereign agentic AI deployments. Tencent also becomes the latest open-source developer to limit its AI cyber capabilities.
Why it matters: Tesla is designing its own AI inference and training accelerators, widening custom-silicon competition.
Tesla detailed custom silicon under new 'Super Intelligence' branding: SI6, an inference accelerator claimed to deliver 80 times more compute, and Dojo D3 training chips that could reach orbital data centers.
Why it matters: Signals rising local resistance to AI datacenter construction across Europe.
Around 100 activists from the Geef Tegengas group broke into a Microsoft data center construction site near Amsterdam, protesting three planned 85-meter towers amid spreading European backlash.
Why it matters: A new always-wearable AI form factor from a high-profile founder tests privacy-first sensing.
Zomato founder Deepinder Goyal revealed Temple, a mic-free AI wearable that uses motion sensors to hear only its wearer, priced around ₹85,000 and shipping in 2027.
Georgia Tech spinout Nagare Labs, backed by Playground Global, showed Z2RAM, a capacitorless memory claiming SRAM speed at DRAM density for AI accelerator stacking.
Nvidia's GPU App Prioritization, which splits GPU resources between apps, first shipped on RTX Spark laptops and now appears in Nvidia App version 11.0.9.585 on regular GeForce RTX desktops, without an announcement.
Why it matters: Export-control enforcement shapes where AI compute can legally flow.
Nvidia is stepping up monitoring of AI accelerators sold in Southeast Asia to stop shipments illegally reaching Chinese firms, after criticism it had not done enough. US Congress is also targeting AI chip exports.
Why it matters: Gives researchers a controlled, reproducible sandbox to diagnose why robot policies fail.
ToyBlocks is an open simulation framework for studying robot learning in a controlled tabletop world: it auto-generates tasks, trains demonstrations and evaluates policies to show how and why they fail.
Cavli Wireless will unveil its P16QS LTE Cat 1bis module and a Cortex AI coding assistant at electronica 2026.
Ray-Ban Meta and Oakley Meta AI glasses are coming to the Philippines with Tagalog AI and four models to choose from. Separately, a woman was secretly filmed at work by someone wearing Meta AI glasses.
China Telecom, Rokid and ZTE launched the first commercial 5G-A and AI-glasses-powered smart tour guide in Taizhou, Zhejiang.
Geekplus's RS Hyper is a warehouse robot that climbs shelves to retrieve totes; RoboShuttle Hyper is rated at 6,000 totes per hour per 1,000 sq m, with a Max workstation handling 800 an hour.
Tower Semiconductor and Innolight said their first jointly developed near-package optics product, built on Tower's PH18M silicon photonics process, has entered early production for AI optical interconnects.
Raycast updated its Screen Awareness feature so users can talk to AI about whatever is on their screen by voice, without explaining context first.
A GitHub blog guide on using Copilot across planning, coding, review and production debugging, arguing the work around code generation still needs human judgment.
A Rust project that provides a headless 3D and VFX engine agents can drive programmatically, letting AI agents render scenes and effects without a GUI. It is a research preview and drew 186 GitHub stars in its first week.
Speechify is an AI reading and voice assistant app. Version 3.24.0 is an update released for macOS, Windows and Linux.
Google's Gemini Notebook gains flashcard tools, with image support among three new features aimed at turning notes into study material.
PyTorch 2.15 ships changes aimed at making large-scale AI training more resilient when a single GPU fails, clearing long-standing technical debt.
Not a Tesla App roundup details which car functions the Grok voice assistant can operate inside Tesla vehicles, from navigation to climate settings.
AWS outlines the four building blocks it converged on while building Amazon Quick: triggers, agents, capabilities and learned context, described as a way to decompose agentic work.
issue-rag is a Python project that answers questions over GitHub issues using retrieval-augmented generation, with a hand-reviewed evaluation set and an audit of the judge model, reporting confidence intervals rather than only best scores.
Saber is a new inference engine claiming 6,645 tok/s prefill and 316 tok/s decode peak for Qwen3.8 27B. Its developer says it takes a different approach from existing engines after a month of testing.
AWS details building a serverless Model Context Protocol server on Amazon ECS Fargate that gives AI agents secure, metadata-driven access to an enterprise data lake; exposing a dataset takes only a JSON file.
Rune is an open-source Rust UI framework that renders interfaces on the GPU and targets macOS, Windows, Linux and the web from one codebase.
Amp, a command-line coding agent and development environment, released two new CLI builds installable via npm, with routine fixes and updates.
A pull request for llama.cpp replaces the SLRU expert cache with a hybrid cache that ranks frequently used experts via decaying usage history. With a warm cache it reports about +6% decode on QFN and +7% on GLM versus SLRU with 8% fewer uploads.
Safegres is a read-only verifier that grades Postgres grants and row-level security in CI, so teams can check database permissions before AI-built apps ship.
Why it matters: Pushes AI coding agents from autocomplete into full app building and deployment.
Cursor's new Build Live Apps feature creates and deploys web applications with databases, authentication and hosting from prompts, TechRepublic reports, detailing what is included and what it costs.
A user wrote their own MCP connector letting Claude drive SolidWorks on their PC to generate custom CAD parts from product listings.
Harness adapters are task- and harness-specific LoRA fine-tunes for a model, letting developers specialize weights for tools like Hermes or OpenCode on jobs such as coding or data analysis.
Box Architect is a natural-language AI tool for building automated workflows around document processes. Announced at BoxWorks 2026, it acts as an agentic operator of Box's content and workflow capabilities.
JuvrixAI launched a no-code platform for building AI agents on NEAR Protocol, letting users assemble and deploy agents without writing code.
Global Times reports ordinary Chinese shoppers are buying and using AI model tokens, pushing the AI boom beyond developers and enterprises into mainstream consumer habits.
Version 1.4.62 of the Grok iPhone and iPad app ships improvements to Chat, Voice and Imagine, the app's image features.
AI Gateway added more zero-data-retention options for Moonshot's Kimi models, letting developers route requests without retaining data.
A free open-source question bank for programmer interviews covering Java, Python, databases, algorithms, LLMs, AI agents and RAG, with company-targeted prep and project walkthroughs; 104 stars in its first week.
A dev tool that runs Claude Code and Codex agents side by side on your subscriptions in parallel worktrees, letting one agent review the other's code and preview app changes.
A free agent app that connects to your existing AI subscriptions, API key or self-hosted model and automatically switches providers when one hits its usage limit; runs on Mac and iPhone.
An all-in-one SEO operating system combining keyword research, AI Overview rank tracking, backlinks, audits and content, plus monitoring of how ChatGPT, Claude, Gemini and Perplexity describe your brand.
A Mac app that answers questions about anything on screen out loud, highlights the real button and walks users through steps; it also builds boards, courses and writes up lectures against slides.
A developer used Claude to build a 15MB Netflix-style front end replacing six separate *arr apps on a home media server, cutting idle memory use.
Why it matters: Shows consumer AI agents taking on real-world phone tasks like scheduling and insurance checks.
Grok's agentic Bot can research nearby specialists, compare options and call offices to book appointments, including handling receptionists and insurance questions, and can also plan family travel.
A Hugging Face Space hosts a step-by-step handbook for training Mixture-of-Experts models with Megatron-LM, aimed at people who have trained dense models and want to try MoE next.
MatterSeek is a Hugging Face Space demo app that lets visitors try the project in the browser; the static demo has drawn over 1,100 likes.
A Hugging Face Gradio Space performs face and head swaps on photos while keeping pose, lighting, garment and composition unchanged; generation runs on a protected backend via ZeroGPU.
A Hugging Face model card describes Qwen3.5-0.8B edited by a frontier LLM through 'Weight-Space Programming' — directly modifying weights with no gradient descent or training — to improve its math ability. Apache-2.0 licensed.
A Hugging Face Space demo that removes clichéd, formulaic phrasing from English and Chinese text while leaving code, URLs, numbers and quotes untouched. Runs fully in the browser with no upload or backend; the Lite build has a fixed rule list.
Nvidia's VFlash method accelerates visual planning by decoding trajectories via flash primitive decoding, reportedly speeding robot trajectory decoding 12x while cutting prediction errors by 61%.
A paper presents WeLM-Image, a unified image foundation model combining text-to-image generation, instruction-guided editing and multi-reference synthesis, built for single-GPU deployment in WeChat consumer apps.
A paper proposes SLA, which builds 3D spatial understanding into vision-language-action foundations via continual pre-training for more reliable robotic manipulation.
A Rust project that provides a headless 3D and VFX engine agents can drive programmatically, letting AI agents render scenes and effects without a GUI. It is a research preview and drew 186 GitHub stars in its first week.
issue-rag is a Python project that answers questions over GitHub issues using retrieval-augmented generation, with a hand-reviewed evaluation set and an audit of the judge model, reporting confidence intervals rather than only best scores.
A free open-source question bank for programmer interviews covering Java, Python, databases, algorithms, LLMs, AI agents and RAG, with company-targeted prep and project walkthroughs; 104 stars in its first week.
An open-source Model Context Protocol server for Roblox Studio that lets AI assistants like Claude and Cursor read object trees, edit scripts and run playtests safely. Written in Python; 425 stars in its first week.
A Python Agent Skill and humanizer package that removes AI-slop, translationese and TDK errors from Turkish text; it drew 106 stars in its first week.
A TypeScript GitHub project that lets you control Claude Code and Codex coding agents from a browser or phone: chats, approvals, terminals, files, git and dev-server previews across your computers, run via one Docker container on your own server.
Ork is a source-available, self-hosted AI assistant whose persistent bots share conversations and work on isolated computers, with web, desktop and mobile apps. The new TypeScript repo drew about 102 stars in its first week.
A new GitHub repo compares long-term memory MCP servers for Claude Code, Codex, Cursor and other MCP clients, covering Glen, Mem0, Supermemory, Zep and others with storage, capture model, team sharing, license and live stars.
A GitHub repo collects 28 SKILL.md templates, agents and MCP configs for AI-powered web security and automation, companion material for an 'Agentic Hacking' ebook; it drew 346 stars in its first week.
Sand3D is a Python library for building 3D models that is designed to be driven by LLMs. The new GitHub repo drew 112 stars in its first week.
agenterm is a self-hosted Go tool letting AI agents control any remote computer via MCP and a terminal; poki runs a local MCP server and agent so AI can operate a real iPhone without jailbreak or extra hardware.
An open-source TypeScript tool that restarts failed AI agent jobs from their last checkpoint instead of from scratch, keeping verified completed work and billing the remainder in USDC on Solana via x402 after you accept results.
Raycast updated its Screen Awareness feature so users can talk to AI about whatever is on their screen by voice, without explaining context first.
Speechify is an AI reading and voice assistant app. Version 3.24.0 is an update released for macOS, Windows and Linux.
Google's Gemini Notebook gains flashcard tools, with image support among three new features aimed at turning notes into study material.
Not a Tesla App roundup details which car functions the Grok voice assistant can operate inside Tesla vehicles, from navigation to climate settings.
Global Times reports ordinary Chinese shoppers are buying and using AI model tokens, pushing the AI boom beyond developers and enterprises into mainstream consumer habits.
Version 1.4.62 of the Grok iPhone and iPad app ships improvements to Chat, Voice and Imagine, the app's image features.
A free agent app that connects to your existing AI subscriptions, API key or self-hosted model and automatically switches providers when one hits its usage limit; runs on Mac and iPhone.
A Mac app that answers questions about anything on screen out loud, highlights the real button and walks users through steps; it also builds boards, courses and writes up lectures against slides.
Grok's agentic Bot can research nearby specialists, compare options and call offices to book appointments, including handling receptionists and insurance questions, and can also plan family travel.
Microsoft Research and Windows Report detail new Copilot accounting skills coming to Microsoft Excel and Cowork this month, aimed at automating accounting workflows inside the apps.
Fathom 3.13.0 is an AI notepad app for macOS, Windows and Linux that records and transcribes video conferences; this update brings bug fixes and performance improvements.
NetEase Youdao launched an interactive math tutoring agent that teaches step by step, takes follow-up questions on any step and draws diagrams on a dynamic whiteboard.
Meta AI's photo expansion tool is rolling out to some WhatsApp users on iOS, including TestFlight and main-app testers, letting them extend the edges of an image in the in-app photo editor.
Granola is an AI-powered notepad that records and summarizes meetings. Version 7.749.1 released for macOS, Windows and Linux.
Perplexity released Teammates, always-on AI agents users can build to handle ongoing work tasks in the background, per The Verge and PCMag.
Superwhisper is a dictation tool that transcribes speech and uses an LLM to reformat the output. Versions 3.2.0 and 3.2.1 ship for macOS, Windows and Linux.
Search Engine Roundtable reports OpenAI updated ChatGPT Ads bidding options to support max bid or max CPC.
Meta AI for iPhone 297.1.0 adds scheduled tasks, daily briefings, email and calendar connections and research reports; separately Meta rolled out Muse voice mode, AI video calls with an animated avatar and custom spaces for US users.
Moonshot's Kimi chat assistant releases version 3.2.25 for macOS, Windows and Linux, an update to the AI chat app; coverage also notes open models like Kimi and Mistral matching frontier rivals.
Canva's Craft AI is now available in India, bringing the design platform's AI creation tools to Indian users with local availability details.
Microsoft is expanding undisclosed usage limits on Copilot Chat for enterprise users without the paid add-on, prompting upgrade offers when features run out.
ElevenReader, an app that reads books, PDFs, docs and articles aloud in natural AI voices, updated its iOS version with source icons and a book assistant that jumps to a location.
An Android app that generates anime pictures entirely on-device, with no cloud, account, ads or root. It runs on Snapdragon 8 Gen 1 NPUs, producing 512x768 images in about 7 seconds.
Apple's Creator Studio now includes new AI features built on OpenAI models, though some paid Logic Pro owners say the new capability is not included for them. The update is now available.
Kokon is a local photo editor built around Qwen-Image 2.1 Turbo. Version 0.12 adds a Layers lab that puts clicked objects on separate layers and repaints the background.
A GitHub blog guide on using Copilot across planning, coding, review and production debugging, arguing the work around code generation still needs human judgment.
PyTorch 2.15 ships changes aimed at making large-scale AI training more resilient when a single GPU fails, clearing long-standing technical debt.
AWS Machine Learning · 5h
#
AWS outlines the four building blocks it converged on while building Amazon Quick: triggers, agents, capabilities and learned context, described as a way to decompose agentic work.
Saber is a new inference engine claiming 6,645 tok/s prefill and 316 tok/s decode peak for Qwen3.8 27B. Its developer says it takes a different approach from existing engines after a month of testing.
AWS Machine Learning · 5h
#
AWS details building a serverless Model Context Protocol server on Amazon ECS Fargate that gives AI agents secure, metadata-driven access to an enterprise data lake; exposing a dataset takes only a JSON file.
Rune is an open-source Rust UI framework that renders interfaces on the GPU and targets macOS, Windows, Linux and the web from one codebase.
Amp, a command-line coding agent and development environment, released two new CLI builds installable via npm, with routine fixes and updates.
A pull request for llama.cpp replaces the SLRU expert cache with a hybrid cache that ranks frequently used experts via decaying usage history. With a warm cache it reports about +6% decode on QFN and +7% on GLM versus SLRU with 8% fewer uploads.
Safegres is a read-only verifier that grades Postgres grants and row-level security in CI, so teams can check database permissions before AI-built apps ship.
Cursor's new Build Live Apps feature creates and deploys web applications with databases, authentication and hosting from prompts, TechRepublic reports, detailing what is included and what it costs.
A user wrote their own MCP connector letting Claude drive SolidWorks on their PC to generate custom CAD parts from product listings.
Harness adapters are task- and harness-specific LoRA fine-tunes for a model, letting developers specialize weights for tools like Hermes or OpenCode on jobs such as coding or data analysis.
Box Architect is a natural-language AI tool for building automated workflows around document processes. Announced at BoxWorks 2026, it acts as an agentic operator of Box's content and workflow capabilities.
JuvrixAI launched a no-code platform for building AI agents on NEAR Protocol, letting users assemble and deploy agents without writing code.
AI Gateway added more zero-data-retention options for Moonshot's Kimi models, letting developers route requests without retaining data.
A dev tool that runs Claude Code and Codex agents side by side on your subscriptions in parallel worktrees, letting one agent review the other's code and preview app changes.
An all-in-one SEO operating system combining keyword research, AI Overview rank tracking, backlinks, audits and content, plus monitoring of how ChatGPT, Claude, Gemini and Perplexity describe your brand.
A developer used Claude to build a 15MB Netflix-style front end replacing six separate *arr apps on a home media server, cutting idle memory use.
Ollama released version 0.40.15 for macOS, Windows and Linux, its tool for getting large language models up and running locally.
Warp released stable version 0.2026.11.04.08.17.stable_01 of its Rust-based terminal for macOS, Windows and Linux.
Swarms v17 InterLink is an open framework release for multi-agent systems that adds six communication channels and a validated fan-out workflow, replacing string passing with a shared conversation. Breaking changes can silently ignore old arguments.
A preview SDK and runtime from Microsoft for building apps that use Copilot, letting developers add Copilot-driven features to their software.
Cursor now lets users talk with their coding agents in a realtime voice call on desktop, web and iOS.
AWS is refitting its Kiro agentic development tools for agent orchestration, adding a bot orchestrator role to address identity questions.
AWS Machine Learning · 1d
#
Web Answer on Amazon Bedrock AgentCore returns a source-cited answer to a natural-language question in one tool call, skipping the search, retrieve and cite pipeline.