Tencent released open-source Hy4 preview model

Tencent unveiled Hy4 preview, a 770B-parameter model with a 1M-token context, built for software, office, gaming, and research work. It outperformed GLM 5.3 and Kimi K3 in internal tests and is now open source.

testingcatalog.com

Exclusive: Deeper look into Hatch Agent from Meta

Meta’s unreleased Hatch materials point to a standalone agent platform with web, iOS, and Android apps, persistent project spaces, connectors, privacy controls, shareable agents, and browser or file-based task execution.

testingcatalog.com

Vellum Mobile brings Voice Mode and Mac control to iOS

Vellum’s mobile app now matches desktop and web, adding on-device voice, camera support, Lock Screen approvals, and remote Mac control. iPhone and Android users can access cloud or self-hosted assistants from anywhere.

testingcatalog.com

Z.ai launches GLM-5.3-Flash under MIT license

Z.ai launched GLM-5.3-Flash, a multimodal 320B-A18B model for coding, agents and visual reasoning. It targets lower cost, 1M-token context, MIT-licensed weights, and reported gains over GLM-5.2 with lower compute and memory use.

testingcatalog.com

Claude unifies memory across Chat and Cowork

Anthropic is launching shared Claude memory across chat and Cowork, with saved context carrying between both. Memory is on by default for most users, editable in settings, and excludes sensitive topics unless users opt in.

testingcatalog.com

Antigravity 2.0 adds new VCS controls and a terminal

Google added Git controls and an embedded terminal to Antigravity 2.0, letting developers review all workspace changes, inspect diffs, run commands, stage files, and commit from the sidebar without leaving the app.

testingcatalog.com

ICYMI: SpaceXAI plans NVIDIA Vera deployment for Starmind

NVIDIA says SpaceXAI will use Vera CPUs for Grok’s agent workloads and plans a space-ready Vera Rubin NVL72 system for the Starmind satellite, extending one AI computing stack from Earth data centers to orbit.

testingcatalog.com

Grok Bot to get Templates and multi-account support soon

TestingCatalog reports xAI is preparing Grok Bot updates including templates, multi-account switching, Chrome profile import, and desktop egress routing, pointing to broader bot setup, account separation, authenticated access, and network control.

testingcatalog.com

Google prepares Gemini App for Avatars, Plugins, and Gemini 4

Google is expanding Gemini desktop with avatars, a hidden Customize hub for apps, skills, and plugins, plus support for richer widgets. The changes also suggest Gemini 4 is in preparation, though rollout timing remains unclear.

testingcatalog.com

Exclusive: Early outputs of Muse Video model from Meta

Meta is beta testing Muse Video with selected partners. Early tests show high-quality 10-second video generation with strong detail and consistency, native audio support, and likely rollout across Meta AI, Instagram, and Facebook.

testingcatalog.com

Meta launches AI desktop app for macOS with screen sharing

Meta has launched a Meta AI macOS app in select regions, adding screen sharing and system-wide dictation to its web features. The release marks a mature desktop step, though coding integration and computer control still trail rivals.

testingcatalog.com

Fabraix opens public access for Nyx AI security red team agent

Fabraix has reopened access to Nyx, an autonomous blackbox red-teaming agent for customer-facing AI. It runs scheduled multi-turn attacks across chat, voice, browser, and coding agents, with replayable findings and CI support.

testingcatalog.com

Ornith-1.5 open models launch in 397B, 35B, and 9 B sizes.

Ornith-1.5 introduces a closed self-improvement loop: it creates harder tasks, builds task-specific scaffolds, and trains on its own rollouts. Across 397B, 35B, and 9B models, it posts strong coding, reasoning, and agent benchmarks.

testingcatalog.com

Google tests Computer Use on Gemini Desktop

Google is testing Gemini desktop computer-use tools that can control apps, access selected files, and share screens. The feature adds backups and oversight controls, but raises concerns around permissions, recovery, and large-folder handling.

testingcatalog.com

Anthropic tests Hub Mode in Claude for sub-agents

Anthropic is testing an internal mobile feature called Hub Mode, possibly linked to a Claude Code hub for managing sub-agents and task boards. Its purpose, functionality, and release timeline remain unconfirmed.

testingcatalog.com

Cursor begins Origin code hosting rollout for paid plans

Cursor has launched Origin in early beta, a built-in code hosting service with repos, pull requests, code browsing, GitHub sync, and agent access, letting teams manage code, reviews, and project tasks from one place.

testingcatalog.com

Vorflux opens sign-ups for its automated coding platform

Vorflux opened public sign-ups for its cloud coding autopilot, taking tasks from intent to merged PR on full running stacks. It uses frontier models, records real test proof, and bills separately for tokens and compute.

testingcatalog.com

Anthropic develops its own Slack as Claude Code Projects

Anthropic’s Claude Code projects now support persistent repo context, shared sessions, and threads. The model points to team-based coding in a maintained workspace, extending Claude from a solo assistant to shared organizational tooling.

testingcatalog.com

Claude turns into Super App, adding web browser for Cowork

Recent Claude Desktop builds point to a built-in browser for Cowork sessions and a user-facing model comparison mode. The unreleased features suggest Anthropic is positioning Claude Desktop as a broader all-in-one AI workspace.

testingcatalog.com

Google launches Sheets canvas for Gemini mini-apps

Google launched Sheets canvas, a Gemini feature that turns spreadsheet data into editable mini-apps built from prompts. The visual layer syncs with the sheet in real time and is available in English for Pro, Ultra, and select Workspace plans.

testingcatalog.com

Google launches Gemini 3.7 Flash for coding and AI agents

Google launched Gemini 3.7 Flash, a faster, lower-cost model for coding, agents and enterprise workflows. It delivers strong benchmark gains, wider product access and updated safety safeguards across consumer and business tools.

testingcatalog.com

OpenAI previews Ultrafast API tier for GPT-5.6 Sol

OpenAI launched a limited API preview of Ultrafast for GPT-5.6 Sol, powered by Cerebras, delivering up to 14x Standard speed and 750 tokens per second for low-latency work across support, finance, security, voice, and research.

testingcatalog.com

ICYMI: xAI releases Grok 4.6 for long-running agent work

xAI’s Grok 4.6 targets long-running agent work, research, and coding, with stronger multi-step execution than 4.5. It posts near-frontier benchmark results, supports sustained project delivery, and is now available across major platforms.

testingcatalog.com

Matic robot adds gesture and voice controls in 75 languages

Matic added voice and pointing control to its cleaning robot, letting users direct spot cleaning without an app. The system processes camera data locally, supports 75 languages, cleans by surface type, and sells for $1,245.

testingcatalog.com

Google tests Agent management UI on AI Studio

Google is adding an agents area to AI Studio for creating, configuring, hosting, and managing managed agents by Google Cloud project. It brings a graphical editor, file tabs, and lifecycle controls atop the existing API-based agent stack.

testingcatalog.com

NoimosAI launches SEO Agent 2.0 for automated optimization

NoimosAI’s SEO Agent 2.0 connects SEO research to site changes, combining Semrush, Search Console, GA4, and audits to propose and, after approval, publish content or fix pages via its builder and WordPress.

testingcatalog.com

Cursor prepares to launch Origin platform for code reviews

Cursor is set to open Origin, internally called Cursor Review, beyond its beta. The platform adds Codebase repo sync from GitHub and an automated PR review flow, marking a near-term rollout as Cursor and SpaceXAI roadmaps converge.

testingcatalog.com

OpenAI plans free ChatGPT Plus year for US college students

OpenAI may offer US students at 220+ campuses a free year of ChatGPT Plus, verified via SheerID. The plan casts ChatGPT as an academic workspace, amid rival student AI offers, but terms and launch timing remain unpublished.

testingcatalog.com

Meta releases Muse Glimmer for local AI agents

Meta released Muse Glimmer, a 30B open-weight model for local always-on agents on consumer GPUs. Apache 2.0 licensed and available on Hugging Face, it supports tool use, multimodal input, 100+ languages, and offline deployment.

testingcatalog.com

Google may retire Gems in October, forcing migration to Skills

Google may retire Gemini Gems on October 20, based on an unreleased in-app notice. The move could shift a free tool to paid, region-limited Skills, with no clear migration path and major concerns for heavy and enterprise users.

testingcatalog.com