From Product Hunt
CLI finance tool, AI coding cost optimizer, and a neat Mac duplicate cleaner with face detection.
Ray — Your Personal CFO in the Terminal
Terminal-based finance tool that reads your real transactions, remembers your goals, and tells you what to do. AES-256 encrypted, data never leaves your machine. Free with your own Anthropic + Plaid keys, $10/mo managed. MIT licensed.
Edgee Codex Compressor — Use Codex at 35% Lower Cost
AI gateway that compresses context sent to OpenAI Codex and Claude Code. Benchmarked: 49.5% fewer input tokens, cache hit rate from 76% to 85%, total session cost down 35.6%. Open source, two commands to install.
Nicelydone MCP — Design Context for AI Agents
MCP server that gives AI agents access to 140,000+ real screens, user flows, and UI components from shipped products. Works with Claude, Cursor, VS Code. Stops AI-generated pages from all looking the same.
Dup v2.0 — Find Duplicate Photos & Remove Faces
Mac app that finds duplicate photos/videos using perceptual hashing (not just filenames). v2.0 adds "Remove Someone" — point at one face, it finds every photo they appear in. Built with Tauri 2 + Rust, all ML on-device. $39.99 lifetime.
ARCH — Cloud Architecture Designer in the Browser
Local-first architecture diagramming for AWS, Azure, GCP, IBM, Oracle, and on-prem. Drag real provider components, nest regions/VPCs, export PNG/SVG/PDF. No account, no server — runs entirely in browser via IndexedDB.
LiquidGlass — Apple Glass Effect on the Web
WebGL library for realistic glass refraction, blur, chromatic aberration, and lighting effects on any HTML element. Supports dynamic content, videos, and draggable panels with per-element configs.
From Reddit
Big week for local AI: llama.cpp gets audio processing, speculative decoding yields huge speedups, and a 0.1B TTS model runs on CPU.
llama-server now supports speech-to-text with Gemma 4 E2A and E4A models. A separate PR also adds Qwen3 Omni audio support (vision + audio input) and Qwen3-ASR. Two major audio PRs landing in the same weekend marks a turning point for local multimodal inference.
Speculative Decoding +50% on Code with Gemma 4 31B
Using Gemma 4 E2B (4.65B) as draft model for Gemma 4 31B on RTX 5090: +29% average speedup, +50% on code generation. TurboQuant KV cache, 128K context. The gains compound because both models share the same architecture family.
MOSS-TTS-Nano — 0.1B TTS That Runs on 4-Core CPU
Open-source multilingual TTS model (Chinese, English, Japanese, Korean, Arabic) with realtime speech generation, streaming inference, and voice cloning. No GPU needed. From MOSI.AI and the OpenMOSS team at Fudan University.
Meta Paper: "Neural Computers" — AI Simulating Terminals & Desktops
Meta trained a video model to generate simulations of terminal and desktop environments. The model learns to produce functional-looking computer interfaces from video data alone. Early research but an interesting direction for AI-driven computer interaction.
Terence Tao: "A Copernican View of Intelligence"
The Fields Medal mathematician argues human intelligence is not the center of all cognition, just as Earth is not the center of the universe. 476 upvotes on r/singularity.
Top Economist: 40% Unemployment and a 3-Day Work Week Are the Same Thing
Fortune analysis reframes the AI labor disruption debate — the question isn't whether jobs disappear, but how work gets redistributed. Whether that's utopian or dystopian depends on policy choices made now.
From Reddit
A game-changing rigging tool for 2D animation, a new image model from Baidu, and useful ComfyUI tools.
Open-source model that takes a single static anime/illustration image and decomposes it into 23+ separate layers ready for rigging and animation. Supports split by side and depth for 30+ layers. Huge for anyone who wants rigged 2D characters without the cost. 659 upvotes on r/StableDiffusion.
ERNIE-Image-8B — New Image Model from Baidu Coming Soon
Baidu preparing an 8B image generation model with ComfyUI and diffusers support already in PR. Early samples suggest strong prompt adherence and photorealism. Watch for the HuggingFace release. 230 upvotes on r/StableDiffusion.
LTX-2.3-VBVR LoRA — Better Video Quality for LTX
Community-trained LoRA for LTX 2.3 image-to-video trained on 96,000 videos. Significantly improves visual quality and motion consistency. Advanced versions with 240K and 480K training data coming soon.
Photopea-Tab for ComfyUI — Seamless Image Editing Integration
Custom node that embeds Photopea directly in ComfyUI's sidebar with bidirectional copy-paste, ad-hiding, and fullscreen support.
From Reddit
An open-source Wispr Flow alternative, a local whiteboard that links to your files, and self-hosted TLS certificate management.
OpenWhisp — Open Source Voice-to-Text for Mac
Hold Fn, speak, release — your words get transcribed by Whisper, polished by Gemma 4 running locally via Ollama, and pasted where your cursor is. No cloud, no account, fully local. A free alternative to Wispr Flow. 161 upvotes on r/SideProject.
VaulTLS 1.1 — Self-Hosted TLS & SSH Certificate Management
Web app for managing mTLS certificates without the complexity of HashiCorp Vault. v1.1 adds SSH certificate management, multiple CAs, and OpenID Connect auth. Built with Rust + Vue.js, runs in Docker.
Whitebloom — Local Whiteboard That Links to Your Files
Unlike cloud whiteboards, Whitebloom lets you drop links to local Excel files, Obsidian boards, and URLs with full cross-linking. No third-party dependencies, you own all your resources. Early but solid. 68 upvotes on r/SideProject.