From Product Hunt

Desktop AI agents that can actually do things, a knowledge graph for your reading, and Playwright finally comes to mobile.

Mindra — Agent Teams You Can Actually Delegate To Command center for agentic teams that collaborate on marketing, supply chain, and ops. Self-healing agents with 3,000+ integrations, compounding memory, and human oversight where it matters.
▴ 284 · mindra.co
Aaavatar — Branded Team Headshots in One Drop macOS-only tool for HR and design teams. Drop photos in, backgrounds removed, colors balanced, missing pieces restored in one click, everyone aligned at same height. Exports every format.
▴ 239 · aaavatar.nl
Manex — Private AI Memory for Documents Upload files, ask grounded questions, preserve useful answers and corrections as persistent memory. Runs locally where supported. Team Brain workspaces without per-seat pricing — $29/mo for a team.
▴ 112 · manex.app
Mobilewright — Playwright for iOS and Android One unified API to automate real devices, simulators, and emulators. Deterministic, auto-waiting, zero-config. Built for developers and AI agents. Free and open source.
▴ 44 · mobilewright.dev
Highlyt — Knowledge Graph for Your Reading Highlights Color-code ideas by meaning, link highlights across books and papers with typed edges (supports, contradicts, expands). AI auto-suggests connections. MCP server plugs your reading graph into Claude.
▴ 18 · highlyt.app
From Reddit

IBM drops a biology model that beats AlphaFold 3 in 9 benchmarks, llama.cpp gets a major speed boost with MTP, and the White House eyes pre-release AI vetting.

Anthropic Co-founder Jack Clark: AI Nearing Automated AI Research ~30% chance by end of 2027, ~60%+ by 2028 that AI research becomes automated with models helping train next-generation models. Argues AI may not need genius-level creativity to self-improve — the strongest evidence is how quickly it moved from coding help to actual research contributions.
Llama.cpp MTP Support Now in Beta — Performance Gaps vs vLLM About to Close Multi-Token Prediction support lands in llama.cpp with Qwen3.5 MTP, more models to follow. Combined with maturing tensor-parallel support, most token generation speed gaps vs vLLM should be erased.
FastDMS — 6.4x KV-Cache Compression Running Faster Than vLLM BF16/FP8 Dynamic Memory Sparsification using learned per-head token eviction achieves massive compression while maintaining quality. Reference implementation from NVIDIA/Warsaw/Edinburgh researchers with reproducible benchmarks.
White House Considers Vetting AI Models Before They Are Released Trump administration exploring pre-release vetting requirements for AI models. A major potential policy shift that could affect open-source model releases and research timelines.
Five Eyes Issue First Coordinated Agentic AI Security Guidance CISA, NCSC, and allies publish guidance prioritizing resilience over productivity for autonomous agents. Threat model shifts from model-level risks to system-level autonomy risks.
LLMSearchIndex — Fully Local Web Search With 200M+ Indexed Pages Python library for local internet-scale search. Custom compressed index of FineWeb + Wikipedia in only ~2GB. Fast retrieval, no paid APIs (Brave, SearXNG) needed. Drop-in replacement for RAG web search.
From Reddit — GenAI

A new realism finetune for the Preview1 base, a mystery open-weights contender beats Flux.2 Pro, and OneTrainer adds Ernie LoRA support.

UltraReal Fine-Tune Anima v1 — Realism Model for Preview1 Base New photorealism fine-tuning for the Preview1 base. ComfyUI workflow included. First version with more planned — generating highly realistic results with strong community reception.
Mystery Open Weights Model Outperforms Flux.2 Pro on Artificial Analysis Unidentified open-weights model appears on the leaderboard outperforming Flux.2 Pro and Z Image Turbo. Community suspects Kandinsky 6.0 from Sber, which was promised for late April release.
Walkyrie-1.3B — Text-to-Image Model Derived from Wan Architecture T2I diffusion model from Wan2.1-T2V-1.3B with pruned ~1B param UMT5 text encoder. Converts the video architecture into an image generation pipeline. Early release.
OneTrainer Now Supports Ernie LoRA Training (16GB/8GB VRAM Presets) OneTrainer adds presets for Baidu's Ernie Image LoRA training on consumer hardware. Full fine-tuning support hopefully coming next.
From Reddit — Apps

Notepad++ fires back at an unauthorized macOS port, a new workspace-restore app for Mac, and a guide for adding RAG to your Paperless documents.

Notepad++ Creator Issues Trademark Claim Against macOS Port The original Notepad++ creator filed a trademark infringement claim against last week's viral "native macOS port." The project is unauthorized and uses the Notepad++ name without permission.
Snapback — Restore Your Entire macOS Workspace in One Click Save a workspace state (apps, windows, positions) and restore it later. Built over 13 months as a native Mac utility. Useful for context-switching between project setups.
Strflow — "Text Yourself" Notes App for Mac, iPhone, iPad Chat-like interface for capturing thoughts — for people who already text themselves quick notes in iMessage or Slack. Now with iPad support, widgets, and a lifetime plan.
n8n + Paperless-ngx + Paperless-GPT — Add RAG to Your Documents Guide combining n8n workflows, Paperless-ngx for storage, and LLM-powered OCR for semantic queries like "How much income tax did I pay last year?" across your scanned document archive.
Speakr v0.8.19 — Self-Hosted Transcription With Prompt Templates Record or upload audio/video, get speaker-labeled transcripts, summarize or chat using your own LLM. New: prompt template variables with {{placeholders}}, auto-export to Obsidian/Logseq. AGPLv3.
BundleHunt Is Back — Discounted Mac Apps Bundle Popular Mac app bundle deal returns with many paid apps at significant discounts. Good time to grab productivity tools.