From Product Hunt
E-ink hardware, a new image gen API, and developer tools worth a look.
reMarkable Paper Pure — The reMarkable 2 successor
$399 monochrome e-ink tablet with 50% faster response, 20% higher contrast, and up to 3 weeks of battery life. E Ink Carta 1300 display, repair-ready design, and integrations with Google Drive, Slack, and Miro. Ships June 2026.
Luma Uni 1.1 API — Reasoning image generation
Multimodal image generation model that interprets intent before generating. Reference-guided generation, multi-reference composition, and culture-aware outputs across styles. ~$0.09 per 2048px image. Top 3 in Image Arena.
Askmeety — 100% local meeting notes on Mac
Everything runs on your Mac: transcription, summaries, next steps, and AI search across meeting history. No bot joins your call, no cloud upload. VisualWalk captures key frames as a blog-style visual summary.
Basedash MCP Server — Your data analyst in every AI tool
MCP server that connects Claude, Cursor, or ChatGPT to your databases, warehouses, and SaaS tools. Pull live numbers, compare cohorts, generate charts. Same access controls your team already uses.
Neo by Amp — Rebuilt CLI coding agent
The Amp CLI rebuilt from scratch for long-running coding agents. Remote control from ampcode.com, automatic context compaction, message queuing, a new Plugin API, and major performance improvements.
From Reddit
OpenAI ships voice models and kills fine-tuning, AMD and a Taiwanese startup bring serious inference hardware to PCIe, and Chrome is quietly installing a 4GB model on your machine.
GPT-Realtime-2 brings GPT-5-class reasoning to voice apps. GPT-Realtime-Translate handles live translation from 70+ input languages into 13 output languages. GPT-Realtime-Whisper streams speech-to-text as the speaker talks.
Pricing: Realtime-2 at $32/$64 per 1M audio tokens, Translate at $0.034/min, Whisper at $0.017/min. Available now in the API.
OpenAI winding down fine-tuning API
Existing customers can create training jobs through January 6, 2027. Inference on fine-tuned models continues until the base model is deprecated. Major shift that forces developers toward prompt engineering and RAG.
AMD Instinct MI350P: CDNA 4 comes to PCIe
144GB HBM3E, 4TB/s bandwidth, 2.3 PFLOPS FP8, standard PCIe Gen5 x16. 600W TDP with optional 450W mode. Targets on-prem AI inference without the thermal density of rack-scale solutions. No GPU-to-GPU Infinity Fabric though.
Skymizer HTX301: 384GB inference card in a single PCIe slot
Taiwanese company packs six chips with 384GB total memory into one PCIe card at ~240W. Runs 700B-parameter models. Decode-first architecture with custom LISA instruction set. Early access available.
Chrome is silently downloading a 4GB Gemini Nano model
No opt-in, no opt-out. Chrome installs Gemini Nano for "Help me write," scam detection, and tab suggestions. Deleting it triggers a re-download. Estimated 6,000-60,000 tonnes CO2 across the install base. Potential GDPR violation.
Malware on HuggingFace: Open-OSS/privacy-filter is an infostealer
Fake "OpenAI privacy filter" model on HuggingFace is actually a Python-based dropper that downloads a PowerShell payload and runs a malicious executable via Task Scheduler. Already reported.
Anthropic: Natural Language Autoencoders — read an LLM's mind
NLAs translate internal model activations into natural language explanations. A verbalizer maps activations to text, a reconstructor maps text back. Used during Opus 4.6 safety audit to diagnose behaviors and detect evaluation awareness.
Multi-Token Prediction for llama.cpp: Gemma 4 gets 40% faster
MTP drafts tokens at 138 tok/s vs 97 tok/s baseline on M5 Max with Gemma 26B. Zero quality loss.
From Reddit
Juggernaut returns for Z-Image, a single-image LoRA technique, and a massive open-source knowledge base for GenAI workflows.
Juggernaut Z — the SDXL classic returns for Z-Image
One of the most popular SDXL fine-tunes is now available for Z-Image base, with FluxKlein 4B and 9B versions in development. Community reception is strong — downloads are already surging.
Single-image style LoRAs via depth map conditioning
New training technique uses depth maps as conditioning to produce flexible style LoRAs from a single image on Klein 9b. Improves character likeness and style consistency without large datasets.
Banodoco Hivemind: 1M+ Discord messages as an agent skill
Three years of discussions from artists and engineers working with open image/video models, packaged as a queryable agent skill. Surface deep knowledge about workflows, techniques, and model behavior.
From Reddit
macOS utilities and a subtitle tool for your arr stack.
Loop — Radial window manager for macOS
Free, open-source window management via a visual radial menu instead of keyboard shortcuts. Hold a trigger key, drag to a zone, release. Supports halves, thirds, quarters, and custom layouts. Native SwiftUI.
Paperman — Make your screen feel like paper
Reduces harsh screen glare by applying a paper-like texture overlay. Unlike Night Shift or f.lux, it doesn't just shift color temperature — it changes the visual feel of the display itself. Lifetime license.
Liqoria — Universal dock music player for macOS
Mini player that lives in your dock and works across Apple Music, Spotify, Deezer, YouTube Music, and more. Album art visualizer, lyrics display, and playback controls without opening the full app.
Contact Sheet — Lightweight photo folder viewer
Open a folder, visually scan everything, find the right image. No imports, no albums, no cloud sync. Lifetime license.
PoRaBar — Menu bar podcast & radio player
Native SwiftUI app. Search thousands of podcasts and radio stations, play from the menu bar. Lifetime license.
Lingarr — Automated subtitle translation for Sonarr/Radarr
Self-hosted subtitle translation that integrates with your arr stack. Multiple translation providers, auto-detection, complement to Bazarr.