From Product Hunt
A Mac prompt manager with a global shortcut, AI-generated explainer videos, AR guidance for physical labor, and a simple outage tracker.
Lunair — Generate Studio-Quality Explainer Videos. Instantly.
Describe your product and Lunair generates a complete explainer video with consistent characters, custom scenes, voiceover, and music. Edit with plain language. Built for landing pages and marketing, not generic AI clips. $100K+ ARR in 2 months.
Prompt Library — All Your Best Prompts. One Shortcut Away.
Save, tag, and organize your AI prompts into collections. Press ⌘⌥P to instantly insert them into any app on Mac. Offline-first, model-agnostic, one-time purchase ($6.35). No subscription.
Upti — All Your Service Status Pages, in One Place
Instant overview of current outages across major cloud providers and SaaS platforms. No accounts, no setup—just check what’s down right now. Free on web, iOS app available.
Doen AI — AI Guidance for Physical Work
Workers wear a camera and AI guides them step-by-step through complex physical tasks—HVAC, manufacturing, etc. Anyone can perform skilled labor immediately without months of training. “The Matrix” for real-world work.
From Reddit
OpenAI acqui-hires the creator of OpenClaw, Google catches a massive model theft attempt, Ant Group drops a 1T-parameter open model, and NVIDIA’s DGX Spark ships impressive hardware with half-baked software.
Sam Altman confirmed that Peter Steinberger, creator of the viral personal agent OpenClaw, is joining OpenAI to drive next-gen personal agents. OpenClaw will transition to a foundation as an open-source project.
“Vibe coded in November, acquired in February”—the fastest AI startup arc yet. Steinberger will lead a new team focused on personal agent infrastructure.
Attackers Prompted Gemini 100K+ Times Trying to Clone It
Google’s Threat Intelligence Group revealed that attackers tried to steal Gemini’s reasoning logic by sending over 100,000 probing prompts in a “distillation attack.” Google detected the pattern, blocked the accounts, and adjusted safeguards. They consider distillation IP theft.
Ling-2.5-1T — Ant Group’s New 1T Open-Source Model
InclusionAI (Ant Group) released Ling-2.5-1T: 1T total params (63B active), hybrid linear attention architecture, 1M token context, MIT license. Claims advantages over DeepSeek V3.2 and Kimi K2.5 in complex reasoning. Pre-trained on 29T tokens.
QED-Nano — How to Train a 4B Model to Prove Hard Theorems
Hugging Face research team published a guide on training QED-Nano, a 4B parameter model that proves olympiad-level theorems in Lean 4. Demonstrates that small, specialized models can outperform much larger general models on formal mathematical reasoning.
DGX Spark Review: Great Hardware, Incomplete Software
Detailed review of the DGX Spark reveals the SM121 software ecosystem is fundamentally incomplete. No official PyTorch CUDA 13.0 ARM64 wheels, many assumptions baked in for x86. 128GB unified memory is impressive but software compatibility issues make it feel like “handheld gaming scraps.”
Quick hits — open models and Apple Silicon tools.
First Time Ever: Top 4 OpenRouter Models Are All Open-Weight
Milestone moment—the four highest-ranked models on OpenRouter are all open-weight. Signals the growing competitiveness of open models.
Qwen3-ASR Now Runs Natively on Apple Silicon via MLX
Ground-up MLX reimplementation. 0.6B model processes a 10s clip in 0.83s (RTF 0.08) on M4 Pro. 4-bit quantized is 4.7x faster. 52 languages.
JoyAI-LLM-Flash — JD.com’s Open-Source Flash Model
JD.com releases JoyAI-LLM-Flash, an open-weight model optimized for speed. Available on HuggingFace.
From Reddit
ComfyUI goes mobile with full workflow editing, and Z-Image Turbo rebrands with new acceleration features.
ComfyUI Mobile Frontend v2.1.0
Major update adds workflow editing on mobile—add/delete nodes, modify connections, reposition elements. Brings full ComfyUI editing capabilities to phones and tablets.
ZPix Adds LoRA Hotswap + SageAttention2 Acceleration
Z-Image Turbo rebranded as ZPix. New LoRA hotswap with automatic trigger word insertion, plus SageAttention2 for faster generation.
From Reddit
A radial pie menu launcher, eMule reborn for macOS, a Screen Studio alternative with 30 new features, and a handful of dev tools and utilities.
Kando — Open-Source Radial Pie Menu for macOS
Cross-platform pie menu launcher that works with mouse, stylus, touch, or controller. Launch apps, simulate shortcuts, open files—all from a customizable radial menu. Free and open-source. Great replacement for LogiOptions+ Actions Ring.
macMule — eMule Revived for macOS
eMule packaged as a self-contained macOS .app via Wine. The ed2k and Kad networks are still alive. Works on Apple Silicon (M1–M4) and Intel. Automated builds via GitHub Actions.
Debut.sh — Screen Studio Alternative with Major Update
30 improvements since launch: multi-display recording, quality presets, recording overlay, smooth zoom/pan, new backgrounds, and more. Native macOS, lifetime license. Screen recording with device frames and motion.
Quick hits — dev tools and utilities.
FrameExporter — Fast Video Frame Export for macOS
Lightweight Swift app for quick frame-by-frame video preview and instant JPEG export. Open source.
Kreuzberg v4.3.0 — Document Text Extraction with Benchmarks
Python document extraction library with interactive benchmarks vs Apache Tika, Docling, and Unstructured.io. Supports PDF, DOCX, images via OCR.
Opencode Manager — Coding Agents on Your Phone
Self-hosted Docker container with Git, file browser, speech-to-text, push notifications. Manage your coding agents remotely from mobile.