From Product Hunt
An AI-coworker army tops the charts, but the keepers are quieter: an open-weight completion model from JetBrains, Slack going all-in on MCP, agent memory over MCP, and a clutch of native Mac toys.
Mellum by JetBrains — Fast open-weight coding models
A family of small, fast LLMs built for low-latency fill-in-the-middle and code completion rather than chat. Mellum2 uses a 12B-total / 2.5B-active MoE (64 experts, Apache 2.0) for roughly 2x the throughput of similar-sized models, with broad language coverage and local-inference support — a credible self-hostable completion model rather than another API wrapper. Weights are on Hugging Face.
Slackbot's MCP Client — 20+ apps, one conversation
Slack's official MCP client turns Slackbot into a connective layer across 20+ partner apps (Atlassian, Linear, Canva, Zoom, Notion, Box, Miro, Replit, Webflow) via a new MCP registry in the Marketplace. Ask in plain language to act across tools — update tickets, sign docs, view dashboards — and run it in a shared channel so the whole team sees what got done. A significant mainstream MCP-adoption signal.
pumaDB — A small hosted memory layer for AI agents
A lightweight memory layer agents read and write through hosted MCP (Streamable HTTP) or a REST API — notes, facts, preferences, task state, transcripts — with scoped tables, per-key rate limits, and automatic version history. Deliberately not a Postgres/vector-store replacement; it's the "give agents a reliable place to remember across sessions" piece. Works with Claude, Codex, ChatGPT and OpenClaw. Tradeoff: hosted, not self-hosted.
WorkClaw — Proactive AI coworkers in Slack & Teams
The #1 launch of the day: multi-agent "AI coworkers" that live in Slack/Teams, each with a job title, a human manager in the org chart, and its own cloud-hosted "ClawOS" computer wired to 3,000+ apps. Agents coordinate with each other over an inter-agent layer and learn skills/routines. Built on the open-source OpenClaw platform with enterprise add-ons (RBAC, credential vaulting, SOC 2). A serious multi-agent-orchestration play; still Early Access and enterprise-leaning.
Reframe — Surf like it's 1999
An open-source macOS browser (Electron/Chromium under the hood) that faithfully recreates the chrome of Safari 1.0, Netscape 4.8, IE5 and Firefox 1.0 — toolbar icons, status bar, spinning throbber — while still rendering today's web fine. A built-in Wayback Mode browses pages as they actually looked back then. Free, fully open source, a delightful native-feeling Mac curio.
Foyer — A room of ambient sound that lives in your notch
A native (non-Electron, battery-light) macOS ambient-sound app where each sound is a glowing orb you place around yourself — distance sets volume, left/right sets pan — using real spatial audio, best with headphones. It folds into the MacBook notch for quick room-switching, with a library of CC0 field recordings (hearth, rain, ocean, Tokyo/Paris ambiences, birdsong). Free for up to 3 sounds per room; a one-time purchase (no subscription) unlocks unlimited.
In The Weights — Does the model know your name?
A clever novelty (built by ex-OpenAI engineers Joey Flynn and Thomas Dimson): type a name and it queries many frontier and small LLMs in parallel, clusters the answers, and scores how strongly the models "recognize" you — max ~996 (Mozart/Shakespeare tier). Flags caveats like hallucinated bios and common-name collisions. A fun probe into model memorization, and a glimpse of what traces you leave as traffic moves off-web into LLMs.
Pixlie — AI video studio with real control
A text-to-video and image-to-video studio pitched on giving you actual control over the output rather than one-shot prompting — shot direction, references and editing instead of slot-machine generation. A generation tool rather than a native app, but worth a look for the GenAI workflow. (Official domain is pixlys.com.)
LocalForge — A local GUI coding agent before code hits git
A free, open-source local coding agent — a sleek desktop workspace (tabs, multi-project, visual diffs) that plugs into any model (OpenAI, Gemini, or local models via Ollama) and runs entirely on your machine inside your repo, so code never leaves it. Positioned as the review/safety pass "before your code hits git history." A natural fit for a Claude Code user who also runs local LLMs.
Inksightful — Search your handwritten diaries
Snap your notebook and journal pages with the iPhone camera and AI handwriting OCR builds a searchable index — find by name, phrase or time, with recognized text paired to the original page image so you can jump back to the source. Library stays on-device; processing uses cloud AI. Currently pre-launch with a beta signup — appealing for bridging analog journals into a searchable digital archive.
From Reddit
A price war breaks out among Chinese labs, the open-weight world builds its own distribution after the Fable takedown, and two small models punch above their weight.
ByteDance, Tencent, MiniMax, Alibaba and Xiaomi all slashed API prices in the same competitive window. ByteDance's Seedance 2.0 Mini dropped to ~$3.40/M, Tencent cut Hy-MT2-Pro ~70%, MiniMax halved its M3 series, Alibaba cut Qwen3.7-Max 50%, and Xiaomi dropped MiMo V2.5 by 99%.
BofA analysts read it as the capability gap between Chinese models narrowing to the point where inference becomes a commodity. Xiaomi's MiMo team pushed back on the "dumping" thesis, saying the new prices are still near break-even.
Cohere North Mini Code — a 30B/3B-active coding model that runs on one H100
Cohere's first developer/coding model: a 30B MoE with only 3B active parameters and Apache 2.0 open weights, built for agentic software-engineering and terminal tasks with a single-H100 footprint. It scores 33.4 on Artificial Analysis' Coding Index, beating Qwen3.5-35B-A3B, Gemma 4 and Devstral Small 2. Available via vLLM/SGLang, the Cohere API and Hugging Face (llama.cpp/Ollama support pending its custom MoE).
Noema Atlas — decentralized P2P distribution for open model weights
An Apache-2.0 desktop app (macOS/Windows/Linux) that distributes LLM weights over a peer-to-peer network built on Iroh (QUIC + NAT traversal), with Hugging Face and mirrors as opt-in fallback. Every file is content-addressed via BLAKE3/SHA-256 and Ed25519-signed, so identical weights from any source dedupe into one verified copy and stream-verify byte-by-byte. Born from the community's reliance on a single US-hosted source after the Fable takedown — a way to rescue and reshare weights that get pulled.
Qwen Code's agentic CLI + VS Code companion is open-sourced
Qwen's agentic coding tool — a CLI plus a VS Code "code companion" extension — is now open on GitHub. Users running it against locally-hosted models via LM Studio report it gets to the point with less tweaking than Continue, Kilo, Cline or Roo, and surfaces context-size and parallel-run config before serving the model. A notably local-model-friendly agentic harness from a major lab.
Paper: data centers modestly lowered US electricity rates, 2015–2024
A new arXiv paper ("Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States") uses an instrumental-variables approach to find data centers caused average US retail electricity rates to fall modestly over 2015–2024 — the opposite of the prevailing narrative. The mechanism: large fixed costs plus economies of scale in transmission, distribution and generation mean durable demand growth spreads costs and lowers unit prices. The authors caution future supply constraints could reverse the effect.
The case for Gemma 4 26B-A4B as the best small model for non-coding work
A well-argued LocalLLaMA thread makes the case that Gemma 4 26B-A4B (a small MoE) is the strongest open model for language learning and health/biology/clinical queries — beating Qwen 3.5/3.6 in those domains even if it trails on coding. Commenters add the dense Gemma 4 12B and 31B variants are also strong for translation, a reminder of how narrow the field of genuinely capable 20–30B models still is.
From Reddit
A new open image model that targets prompt-following, the LTX video node grows a full editing environment, and a surgical way to edit with Ideogram 4 without regenerating.
A new open-weight image family dropped under Apache 2.0 in Base, Turbo and Edit variants. It's a 10B model that uses the full Qwen3-VL-8B as text encoder and reuses the FLUX.1 VAE, with benchmarks claiming it edges out larger models like Qwen-Image-2512 (20B) and Hunyuan-Image-3.0 (80B) on prompt-following.
One tester put it through complex UI/app-screen generation and found instruction-following strong; Turbo runs at CFG 1 in 4–8 steps, and INT8 quantization (via the ComfyUI-INT8-Fast node) speeds it up on RTX 30-series cards. Repos and fp8 builds are on Hugging Face and GitHub.
LTX Director 2.0 — a free, all-in-one AI video editing node for ComfyUI
A complete overhaul turns local LTX 2.3 video generation into an editing environment inside one ComfyUI node: trim/split/combine on a timeline, video extension via prompts + keyframes + audio, IC-LoRA support (drag-and-drop a video onto an IC-LoRA track), Retake Mode and audio inpainting — bundling t2v, i2v, first/last/middle-frame control, Prompt Relay and custom audio. Example workflows ship in the repo.
Surgical edits with Ideogram 4 via SAM2 masks + partial denoise
A detailed, workflow-included technique for targeted edits: instead of regenerating, it uses SAM2 to auto-generate masks from a simple text prompt ("face, hair" or "background"), then does img2img inpainting at ~0.90 partial denoise so the model keeps the body/scene structure. Captioning and a structured JSON prompt are composed by any LLM (the author runs a local Gemma 31B), so you change only the targeted object while preserving identity. ~2 min on a 3090 stack; workable on a single 12GB card with an online LLM for prompting.
From Reddit
A generous haul this round: native Mac utilities that reclaim hot corners and system data, a local-first notes app, and three self-hosted projects worth a Docker slot.
Safebucket v0.6.0 — self-hosted, S3-agnostic file sharing
A simple open-source alternative to WeTransfer/Dropbox for sharing files from your own S3-compatible storage. v0.6.0 adds in-UI file preview (images, video, text), bulk multi-file download, and LDAP authentication. Docs at docs.safebucket.io.
Fluxer — self-hostable, Discord-style community chat
An open-source community chat app working toward federation. A major update brings all development back to public repos with fully supported self-hosting (docs + prebuilt Docker images, hundreds of reliability fixes) plus new mobile clients — Android is live, iOS is in closed beta. A real self-hosted alternative for teams that want to own their chat.
trinket.io was shutting down — so they open-sourced it
trinket.io (the in-browser Python/code learning sandbox) was being shut down and open-sourced itself instead. A community member now hosts it free at trinket.strivemath.org, and the upstream OSS repo lets you self-host your own instance — a nice preservation-of-software story for educators and tinkerers.
CCCCorners — make macOS hot corners actually useful
A $2.99 one-time utility that lets you assign app launches, keyboard shortcuts and automations to screen corners across multiple displays, now with presets, automations and zones. A lightweight, focused alternative to BetterTouchTool's corner triggers. macOS 11+.
DevKnife — 30 offline developer tools in one native Mac app
A native macOS app bundling 30 offline dev tools (Base64, JSON inspector, HTTP tester, IP location, WHOIS, port scanner, hashing, ruler) plus syntax-highlighted Quick Look previews. Privacy-focused and fully offline — a native alternative to DevUtils/DevToys. $29 one-time with a 30-day trial; just hit v1.15.
RetroMac — classic OS themes with authentic CRT shaders
From the maker of Reframe (above): an app that transforms your whole desktop into a retro machine with full OS themes (Windows 98/XP, Mac OS 9, Mac OS X, BeOS, Snow Leopard) plus CRT glow, VHS flicker and Game Boy shaders, applied per-screen, per-window, even to the webcam. Themes and 15+ shaders are free; €8.88 lifetime unlocks extras. Open source on GitHub.
Mac SysData Cleaner — tackle the "System Data" black hole
A free, open-source utility for clearing macOS's mysterious "System Data" storage bucket. It routes everything to Trash first (so nothing is permanently nuked) and offers one-click Google lookup for unknown files. A free alternative to paid cleaners; .dmg from GitHub releases.
Glyph — a completely local, open-source Markdown notes app
An Obsidian-style notes app (Tauri 2 + Rust) where notes stay as plain Markdown files with wikilinks, backlinks, a local graph view and built-in AI tools — no cloud, no phone-home. For the privacy-minded; community builds are free.
Moodesk — per-Space rules for your Dock and menu bar
A menu-bar app that attaches rules to each macOS Space, so switching desktops can toggle Dock visibility/position, the menu bar, Dark Mode or Stage Manager — or run shell scripts, AppleScripts and Shortcuts automatically. Niche power-user automation; macOS 15+.
Nefoin — auto-install any Nerd Font in seconds
A small CLI that installs any Nerd Font without manual downloading or repo cloning — handy when setting up a new terminal or dev environment. Open source.
fli — a tiny, readable file-listing tool
An 18KB (on a Pi Zero)
ls companion written in Rust no_std + libc that uses emojis instead of color coding for readability, with short/long and sorted output modes. Built for accessibility and minimal SSH environments.