From Product Hunt
AI agent tooling dominates today — video editors, generative UI standards, and Claude Code memory tools. Plus a solid macOS text expander update.
Cardboard — Agentic video editor
YC W26-backed AI video editor. Describe edits in natural language, AI proposes cuts on timeline. Browser-based with multi-track editing, shot detection, beat sync, voiceover generation, and voice cloning. Exports to Premiere/DaVinci/FCP XML.
OpenUI — Open standard for Generative UI
LLMs respond with interactive UI components (cards, tables, forms, charts) instead of plain text. 67% fewer tokens than JSON, 3x faster rendering. Model-agnostic and framework-agnostic. Open source.
Gemini Embedding 2 — First natively multimodal embedding model
Maps text, images, video, audio, and documents into a single 3072-dim embedding space. Native audio embeddings without transcription. 100+ languages. Matryoshka representation learning for flexible dimensions.
Firecrawl CLI — Web data toolkit for AI agents
Scrape, search, crawl, and browse the web with results written to filesystem. 96% web coverage including JS-heavy pages. Works with Claude Code, Codex, OpenCode. 91K GitHub stars.
TADA — Open-source TTS with zero hallucinations
Speech-language model with 1:1 text-acoustic token alignment. 5x faster than conventional LLM-based TTS. Zero hallucinations by construction. 1B English + 3B multilingual models, designed for on-device deployment.
CodeYam CLI & Memory — Memory management for Claude Code
Background agent reviews coding session transcripts, identifies confusion patterns, and generates targeted Claude Rules with path scoping. Prevents repeated mistakes and stale CLAUDE.md files. Free, local, no login.
ELVES — Multi-agent workspace for Claude Code & Codex
Each task in its own git worktree. SQLite-backed memory with relevance decay. Skills and MCP server management via UI. Telemetry dashboards for token usage and session timelines. All local, MIT licensed.
Typinator 10 — macOS text expander, now with iOS companion
Major update with new iOS companion app. Multi-device sync (cloud-agnostic — iCloud, NAS, Dropbox). Apple Intelligence integration for creating and translating expansions. One-time purchase ($39.99).
From Reddit
Anthropic's recursive self-improvement is no longer theoretical. M5 Max ships with stunning local inference numbers. Nvidia commits $26B to open-weight models.
Behemoth TIME profile reveals recursive self-improvement is "a present phenomenon." Anthropic revenue nearly doubled to $20B annualized pace. Enterprise share hit 20% (from 4% a year ago) while OpenAI fell from 50% to 27%.
Fully automated AI research could be roughly one year away. The piece frames Anthropic as the most disruptive company in Silicon Valley right now.
Apple M5 Max Ships Today — First LLM Benchmarks
Qwen3.5-122B-A10B 4-bit: 881 t/s prompt processing, 65.9 t/s generation at 4K context on M5 Max 128GB. Massive upgrade for local inference. Community benchmarking in full swing.
Nvidia Disclosed $26B Investment in Open-Weight AI Models
Largest single commitment to open-weight model development in AI history. Strategy: commoditize your product's complement — free models keep CUDA as the default inference target.
Nemotron 3 Super — 120B MoE Hybrid Mamba-Transformer
12B active params, 5x throughput vs previous. 1M token context via Mamba-2 linear-time layers. Pretrained on 25T tokens in NVFP4 (Blackwell-optimized). Fully open: weights, datasets, and recipes.
Perplexity's "Personal Computer" — Always-On Mac Agent
AI agent running on Mac mini with local file and app access, AI processing on Perplexity servers. Controllable from any device. Perplexity Max subscribers only ($200/mo), waitlist.
llama.cpp Adds Reasoning Budget Support
Counts thinking tokens and forces termination at budget. Prevents reasoning models from spiraling into 2000-token think loops on simple questions.
From Reddit
ComfyUI's Anima model gets a major upgrade, LTX Desktop adds Linux support, and Xiaohongshu drops a competitive open-source image editor.
Anima Preview 2 — ComfyUI's In-House Model Gets Major Upgrade
2B text-to-image model from Comfy Org. Preview 2 brings better natural language comprehension, improved finetuning robustness, and longer training for character knowledge. Community reports significant improvement in creativity and steerability.
LTX Desktop — Linux Support Coming, Community Fork Hits 6GB VRAM
Update: Official team shipping Linux support and IC-LoRA integration. Community's WanGP fork reduces VRAM requirement from 32GB to 6GB. Adding Claude Code skills for contributor alignment.
FireRed Image Edit 1.1 — Open-Source SOTA Image Editing
Xiaohongshu's editing model based on Qwen Image. SOTA identity consistency, 10+ element fusion, comprehensive portrait makeup. Results rival closed-source models. Free ComfyUI workflow available.
From Reddit
macOS utilities and open-source updates — a Tahoe-native clipboard manager, Apple Intelligence as a local API server, and OBS gets a big streaming upgrade.
ClipDoc — macOS 26 Clipboard Manager
Built exclusively for macOS Tahoe. OCR for images, file copy/move from Finder, preview/save/share, drag-as-files. Freemium with $2.99 one-time for unlimited storage.
Apple Intelligence as an OpenAI-Compatible API Server
SwiftUI app turns Apple Intelligence into a local OpenAI-compatible API. Point any client at localhost:11535 — no API key, fully on-device. Runs as GUI app to avoid Apple's CLI rate limits. Free, open source.
DriveVault — Offline Hard Drive Cataloging for macOS
Scan external drives once, search forever. Visual previews, metadata capture, drive comparison. Modern alternative to CDFinder/NeoFinder.
OBS Studio 32.1.0 Adds WebRTC Simulcast
Stream multiple quality tiers without server-side transcoding. Cheaper self-hosted streaming with E2E encryption support.
devPad — Slide-Out Dev Panel for macOS
Split browser, code editor, AI chat, and 14 built-in developer tools in one slide-out panel. Lifetime license.