From Product Hunt

AI phone agents, server management, and a macOS focus app. Also a strong v4 release from imgproxy for self-hosters.

PollyReach — AI Phone Agent for Everyone Give your AI agent a real phone number. It makes calls, handles conversations naturally, navigates IVR menus, waits on hold, and reports back with summaries and transcripts. Also answers your phone 24/7 and screens spam. Works in 50+ languages.
▲ 432 · pollyreach.ai
CtrlOps — AI-Powered Linux Server Management AI terminal that generates commands with your approval, one-click deploys from GitHub, visual file manager, and real-time monitoring. 100% local—your credentials never leave your machine. Mac, Windows, Linux.
▲ 213 · ctrlops.io
Motion — Video Agent for Motion Design Give it a prompt with links, assets, or references. It researches, storyboards, and creates explainers, launch videos, and logo animations. Edit everything directly at the element level—no regenerating entire scenes.
▲ 166 · motion.so
Monocle 3.5 — Noise-Cancelling for Your Screen Dims everything except the active window with a cursor shake. v3.5 adds App Groups (keep related apps visible together), Stage Manager support, fixed multi-monitor blur, and Corner Peek. 50% off until May 31.
▲ 144 · heyiam.dk
imgproxy v4 — Self-Hosted Image Processing Fast, secure, self-hosted image processing server. v4 brings parallel downloading, internal caching, RAW format support, and image classification. Used by Medium, Dribbble, and Substack. Single Docker image.
▲ 134 · imgproxy.net
Haystack — AI PR Review for Agent-Generated Code Triages each PR before you read diffs: safe to merge, needs fixes, or needs human review. Analyzes diffs, codebase context, and the agent conversation that produced the PR. Built for the 20+ PRs/day era.
▲ 114 · haystackeditor.com
From Reddit

Massive day. Google I/O dropped Gemini 3.5 Flash, Antigravity 2.0, and Gemini Omni. Karpathy left OpenAI orbit for Anthropic. Intel leaked a 160GB GPU that sidesteps HBM shortages.

Google I/O: Gemini 3.5 Flash + Antigravity 2.0 Gemini 3.5 Flash is Google's strongest model yet for coding and agentic tasks—4x faster than other frontier models. Antigravity 2.0 is a new agent-first desktop IDE. Demo: 96 agents built a working OS from scratch in 12 hours for under $1K. Pricing is 3x more than 3.2 Flash though.
Gemini Omni: Google's Unified Video Generation Model Replaces the standalone Veo line. Generates and edits video from text, images, video, and audio inputs with conversational editing. Omni Flash available now for AI Plus subscribers. Coming to YouTube Shorts/Create this week.
Cerebras Runs Kimi K2.6 (1T Params) at 981 Tokens/s Cerebras achieved near-1,000 tok/s on Moonshot AI's trillion-parameter open-weight model—6.7x faster than the next-fastest GPU cloud, 23x faster than median inference providers. Enterprise trials now available.
Intel Crescent Island GPU Leak: 160GB LPDDR5X, No HBM Leaked PCB shows Intel's Xe3P data center GPU with 20 LPDDR5X modules for 160GB total, bypassing the HBM shortage entirely. 640-bit memory interface, estimated 704-760 GB/s bandwidth at 8800-9500MT.
Nvidia Nemotron-Labs-Diffusion: Tri-Mode LLM Switches between autoregressive, diffusion, and self-speculation decoding in a single architecture. 8B variant produces 5.9x more tokens per forward pass than Qwen3-8B at same accuracy. 850-1,015 tok/s on GB200. 3B/8B/14B variants, open weights.

More from the local AI and research scene.

HuggingFace Carbon: Open DNA Foundation Models 3B model matches SOTA Evo2-7B while being 275x faster. Trained from scratch using SmolLM methodology adapted for genomic data.
HRM-Text 1B: $1K to Train, Beats Llama 3.2 3B on MATH Hierarchical reasoning architecture. 1B params trained on 40B tokens in 1.9 days on 16 GPUs. Benchmarks look suspicious but the approach is novel.
From Reddit

LTX 2.3 ecosystem keeps growing. New RL-LoRA for audio+video sync, RTX upscaler node, and SDR-to-HDR conversion.

LTX 2.3 OmniNFT RL-LoRA: Synced Video + Audio RL-trained LoRA for LTX 2.3 that greatly improves video quality, reduces artifacts, and achieves perfectly synchronized audio. Realistic lip-sync, action-matched sound, 52% fewer sync errors. Released by Kijai.
Nvidia RTX 2-Pass Upscaler for ComfyUI Custom ComfyUI node implementing 4 NVIDIA RTX upscaling methods. Runs on just 4GB VRAM + 8GB RAM. Ideal for upscaling AI-generated video.
LumiPic: SDR to HDR Conversion LoRA Converts standard images to float-valued HDR EXR files. Even if you don't need HDR, the extra range helps with exposure and color editing. ComfyUI workflows included.
From Reddit

Plex just tripled its lifetime pass to $750. Plenty of new macOS apps: a native Google Calendar client, a batch photo editor, screen recorder, and more.

ScreenKite 1.5.0 — Native macOS Screen Recorder Claims 3x faster than Screen Studio. Dynamic layout for short-form video, AI editing integration with Claude Code/Codex/Antigravity. Native macOS, one-time purchase, no Electron.
Native Google Calendar App for Mac (TestFlight) SwiftUI-built Google Calendar client with real-time sync, Meet links, menu bar widget, and keyboard shortcuts. 265+ testers, App Store launch planned for June. Fills the gap between Apple Calendar's poor Google support and Fantastical's $60/yr.
RapidPhoto — Batch Photo Editor for Mac Edit one photo, apply to all 499 others. Fills the space between Lightroom's $12/mo subscription and PhotoBulk's limited editing. One-time purchase on Mac App Store.
Zipora — Native macOS Archive Tool in Swift/SwiftUI Compression, extraction, preview, extraction history, and archive modification. One-time purchase.
DockFolders — iPad-Style App Folders for macOS Dock Group apps, files, and shortcuts into customizable Dock folders. 50% off right now.
PrimeTask — Local-First Productivity Workspace for Mac Tasks, projects, CRM, notes, visual planning. Data as plain JSON, works offline, optional BYO AI. One-time purchase, Obsidian integration.
Muxy — Terminal with Mobile Remote Control Built on libghostty with Git integration, project sidebar, file tree, and a companion iOS app to manage sessions from your phone. Open source.