From Product Hunt
Hardware, local AI, and developer tools lead today. SpeakON is a neat physical device, Cai brings local AI to a single keystroke, and Cavalry going free is a big deal for motion designers.
SpeakON — A MagSafe AI device for a post-keyboard world
MagSafe hardware device for iPhone — press once to speak into any app. No mic permission needed, works with your phone locked. Cleans up speech automatically and adjusts tone per destination.
InstantDB — Complete backend with auth and storage in one prompt
Full-stack backend with real-time sync, auth, permissions, and storage. 100% open source, built on Postgres. Endorsed by a Firebase co-founder — works exceptionally well with AI coding agents like Codex.
Cai — Press ⌥C on anything to run smart actions, locally
macOS app — one keystroke for AI prompts, shell scripts, GitHub issues, Linear tickets. Bundled Ministral 3B for zero-setup local AI, or connect Ollama, LM Studio, Apple Intelligence. Free, MIT licensed, no telemetry.
Cavalry Studio — Free Motion Design tool by Canva
Professional 2D animation software now completely free via Canva. Character rigging, rubber hose animation, data-driven workflows, and Lottie export. Mac and Windows.
VibeAround — Chat with your local AI coding agent from any IM
Tauri/Rust desktop app connecting Telegram, Slack, Discord to local AI agents like Claude Code, Gemini CLI, Codex CLI. Handover sessions between terminal and phone with full context.
Loomal — Identity infrastructure for AI agents
Gives every agent a DKIM-signed inbox, encrypted vault, and per-action 2FA. One MCP server with 18 tools. Works with LangChain, CrewAI, Claude, OpenAI, Cursor, and any MCP client.
Reference Board — Infinite Canvas for your ideas
Native mood board app for iPhone, iPad, and Mac. Infinite canvas with iCloud sync, Apple Intelligence-powered search, auto-tagging. No subscription, private by design.
From Reddit
Massive day for local models. Qwen's new 27B dense model punches way above its weight class. Meanwhile, GPT-5.5 briefly leaked, Google unveiled new TPUs, and Mythos keeps finding bugs.
Alibaba's latest dense model surpasses Qwen3.5-397B-A17B across all major coding benchmarks despite being 15x smaller. Apache 2.0, supports thinking and non-thinking modes, native multimodal.
Community reports it's usable with Claude Code on RTX 5090 at 50 tok/s with 200K context. The best local coding model at this size class — and the dense vs. MoE gap is shrinking fast.
Update: Mythos finds 271 bugs in Firefox — 12x more than Opus 4.6
Following March's audit where Opus 4.6 found 22 vulnerabilities, Mozilla used Anthropic's Mythos Preview on Firefox 150 and found 271 bugs in a single pass. CTO Bobby Holley confirmed the 12x improvement in scale and speed.
Google unveils TPU 8t and TPU 8i — 8th-gen AI chips
TPU 8t for training reduces frontier model development from months to weeks (2.8x price/performance). TPU 8i for inference delivers 80% better performance-per-dollar with 3x SRAM. Both scale to 9,600+ chips per cluster. Shipping later 2026.
OpenAI accidentally exposes GPT-5.5, Arcanine, and Glacier in Codex
A routing misconfiguration briefly exposed unreleased models in the Codex dropdown for 47 minutes. GPT-5.5 and oai-2.1 tagged as "latest frontier agentic coding model." Internal codenames Arcanine and Glacier-alpha also spotted.
Tencent and Alibaba in talks to invest in DeepSeek at $20B+ valuation
Chinese tech giants negotiating stakes in DeepSeek, which doubled its funding target due to exceptional demand. The startup that built V3 for ~$5.6M is now commanding a $20B+ valuation.
Quick hits — new models and research.
Update: Xiaomi MiMo-V2.5-Pro — 1,000+ tool calls per task
Public beta with massive agentic improvements. Built a full Rust compiler in 4.3 hours, completed an 8,192-line desktop app in one session.
OpenAI open-sources privacy-filter model — Apache 2.0
1.5B parameter PII detection model (50M active). Detects 8 categories at 96% F1, runs locally on laptops.
Sony's AI robot beats pro table tennis players — Nature cover
Project Ace defeated world top-25 players using event-based vision and reinforcement learning. First autonomous robot competitive at elite level in a physical sport.
From Reddit
3D generation gets more accessible, and local voice synthesis keeps getting better.
Trellis.2 optimized for 8GB GPUs — insane 3D detail from images
Community fork brings Microsoft's 3D generation model down from 24GB to 8GB VRAM with a single-click installer. 1024² voxel detail, RTX 3060 completes in 13 minutes. A1111-style ease of use.
Qwen3 TTS — seriously underrated real-time local voice synthesis
One of the most expressive open TTS models available. Sub-97ms first-packet latency, voice cloning, runs fully local. Community's Persona Engine adds Live2D avatar integration for real-time AI conversations.
Local manga translator with built-in LLM, written in Rust
Combines object detection, visual LLM-based OCR, layout analysis, and fine-tuned inpainting. Integrated llama.cpp, supports Gemma 4 and Qwen. Fully offline.
From Reddit
Mac utilities — a QuickLook upgrade for devs, a native Twitch client, and better dock gestures.
FinderPeek — Bundle of QuickLook plugins for developers
Adds QuickLook extensions for 20+ dev file types. Syntax highlighting, interactive JSON viewer, tabular data views, markdown preview. $4.99 one-time purchase.
Kulve — Fully native Twitch app for macOS
Native macOS Twitch client with smooth, lag-free playback and minimal CPU usage. No Electron. Currently in public TestFlight beta.
Click2Minimize v3.9 — macOS dock and window gestures
30+ mouse and trackpad gestures for dock and window management. Complementary to BetterTouchTool.