0830 | Tencent Hy4 preview: open 770B MoE, 49B active, 1M context; Cohere Parse 5

Mia: Welcome to Product Hunt Daily on Bri Radio. I'm Mia.

Milo: And I'm Milo. Today we're diving into some fresh launches, including Tencent's big open-weights model, and a few handy tools that caught our eye.

Mia: We've also got a document parsing model from Cohere, a new way to review code diffs, and several clever Mac and Android utilities worth a look.

Milo: Plus an AI community built around one simple travel question, an investor's pitch deck analyzer, and even a browser-based spy satellite simulator. Let's get into it.

Milo: Tencent has put out the Hy4 preview — an open mixture-of-experts model with 770 billion total parameters, 49 billion active, and a 1-million-token context window, framed by Tencent for long-horizon agentic work: autonomously coding, doing game development, and complex document analysis, running its own tests and fixing bugs before delivery. The launch write-up calls it the step after Hy3, with training built substantially around real work from Tencent's own engineering and specialist teams, so it stays with a task when that task gets long or messy.

Mia: The agentic angle maps onto a known failure mode. One commenter notes the model overthinks and over-checks itself — the same habit as Hy3, release it and see what breaks — so on long jobs the real risk is burning the context window re-checking things already settled, and the 1-million-token room makes that behavior easier to do and harder to spot. That same commenter argues the overthinking is what to watch rather than the eval numbers.

Milo: Those numbers are worth reading carefully. Tencent ran an internal eval with 163 of its experts across 203 engineering tasks, and Hy4 preview came out slightly ahead of GLM 5.3 and Kimi K3 — and because it's an internal blind eval, it's not an independent benchmark. The model is open under Apache 2.0 and available now through OpenRouter or WorkBuddy. And on deployment, another commenter points out that 49 billion active parameters decides whether anyone can run it outside a datacenter — but the 770 billion total parameters still have to be parked somewhere.

Milo: Cohere Parse 5 is Cohere's document vision parsing model — pitched on Product Hunt at enterprise documents that are messy and unstructured, full of tables, diagrams, and charts. Cohere's stated problem is that most AI agents and search tools can't reliably read those, so companies end up doing manual review and data entry. The differentiator it claims is that Parse combines OCR with multimodal understanding: it doesn't just extract text, it understands layout, tables, and diagrams together.

Mia: What makes that interesting is the visual grounding. Every extracted piece of data gets bounding boxes, traceable back to its exact location on the page — which is what makes citations and source attribution possible downstream in an AI pipeline, rather than just pulling text out of a scan.

Mia: Cohere lists OCR for scanned and digital documents, multimodal parsing of tables and embedded images, visual grounding, support for nine major commercial languages, and deployment across its API, Model Vault, AWS SageMaker, Azure, or fully on-premises air-gapped environments. It's positioned for high-volume enterprise documents like claims, contracts, invoices, and reports, feeding internal AI search, retrieval-augmented generation pipelines, or multimodal agents. The target use cases are automating claims, contract, and invoice processing, and improving the quality of chunking and retrieval. These are all Cohere's claims, not independently verified results.

Milo: seendiff is a code diff viewer with progress tracking, launched on Product Hunt by its maker as a local diff tool for large changes. His framing is that coding agents can generate huge diffs and someone still has to read them, and that most diff tools were built for the pre-AI world where code didn't come in that scale — so this was built to scratch his own itch.

Mia: The core feature is automatic tracking of what you've already seen in a diff, plus tracking of manually reviewed blocks, a walkthrough mode where your AI walks you through the code, and local global search with a minimap. It's free, open source, and available at seendiff.com.

Mia: In the discussion, one commenter said they lose track of what they already reviewed, and another called the seen-tracking what would get them using it daily — then asked whether that state is local to one machine or syncs if you pull the same branch on another machine. A third commenter liked seen-state as the right primitive but wanted a why-state distinction, because a hunk skimmed for two seconds and one actually reasoned about are both marked seen, and they're not the same risk on a nine-hundred-line agent diff. So there's an open question: whether seen-state syncs across machines was never answered in the thread.

Milo: Einfall is an app for capturing fugitive thoughts and routing them where they belong — on Mac, iPhone, and iPad. You grab a thought from the menu bar, share sheet, Shortcuts, widgets, Spotlight, or Siri, then process it by sending it to Reminders, Calendar, Files, a Shortcut, or giving your AI agent control over it through MCP.

Mia: So it's less a notes app and more a routing layer — the capture is free and unlimited, but after the first fifteen routing actions there's a one-time 30 dollar purchase to unlock unlimited routing. The pitch is that the thought shows up at the strangest times, and this catches it before it's gone.

Mia: There's a new Mac app called Mossy, and its makers describe it as a desk plant on your desktop that slowly wilts the longer you go without a break. You write one sentence about what a break should be for you — like stretch my bad back, twenty push-ups, or just make me stand up — and an AI turns that into one task at a time, delivered on a timer.

Milo: The personal story behind it explains the design. The developer says he sits about ten hours a day, has dealt with lower back trouble for years, tore his ACL last October, and has spent the year in surgery and physio. He and his wife, an architect, built Mossy for themselves after deleting every break app they tried, because full-screen breaks landed mid-sentence, got skipped, and were gone within a week. That's why it never blocks or dims your screen, and the page is explicit that ignoring it is a normal way to use it, not a failure.

Mia: And the privacy side follows the same restraint. No account is required, and the makers say your history, garden, and settings all stay on the Mac. The only things that leave it are the AI request containing your programme and your last three notes, plus an anonymous daily ping you can switch off, and the AI key lives on a small relay — never inside the app. They also state plainly that it's not medical advice. It's free, native macOS, runs on macOS twelve and later across Apple Silicon and Intel, and comes in at just one point four megabytes. The app counts your keyboard and trackpad use and freezes when you step away, shows one task in steps with a countdown, and grows a sprig each break so you can see the week at a glance.

Mia: Another Windows-facing tool here, but in the other direction. Any Command is an Android app that turns your phone into a second screen, trackpad, keyboard, and gamepad for a Windows machine. The maker built it because he has a single monitor and didn't want to buy a second one — alt-tabbing out of fullscreen kept killing whatever was playing, so it streams just one PC window to the phone, and input from the phone goes only to that window. That way a film stays fullscreen on the monitor while you work from the phone, over Wi-Fi on your own network, from anywhere over the internet, or over Bluetooth with nothing installed on the PC.

Milo: It's grown over about two years beyond remote control. There's now a trackpad with gestures, a keyboard, custom shortcut panels, a home-screen widget for automations, file transfer, clipboard sync, a gamepad, and live CPU and RAM views. Remote control, video remote, Wake-on-LAN, the performance monitor, and clipboard sync are free, while screen share, remote split, file transfer, the virtual gamepad, automations, second screen, and program launcher are premium — and every premium feature can be trialled temporarily by watching a short ad. The maker says there's a free tier and no account, and the PC side is just a plain Windows ten and eleven server install.

Mia: There's one mode worth knowing for locked-down environments. Bluetooth pairs the phone as an ordinary Bluetooth keyboard and mouse, and the maker says that's the only mode that works on those machines. In the comments, one user called the fullscreen problem a specific, real annoyance and said they'd put a live terminal or build log on their phone.

Milo: Switching to the Apple side, there's a utility called RawToHEIC, built by Sandro Gumz, that converts camera RAW and ProRAW files to HEIC straight from the Photos share sheet on iPhone, iPad, and Mac. The maker says converted files can be up to ten times smaller, with edits, albums, and metadata intact — everything runs on-device, with no uploads, no account, and no analytics. It's aimed at photographers shooting RAW with mirrorless or DSLR cameras, or iPhone ProRAW, who watch their Photos library swell into tens of gigabytes. The app shows how much space would be recovered before you convert, and originals only get deleted after the HEIC is written and verified inside the library, and it works with iCloud Photos.

Mia: The maker gives a concrete scale for this — a two-thousand-five-hundred-photo library using more than eighty gigabytes. Pricing is free for fifty conversions a day, and unlimited use is either a one-time nineteen-ninety-nine purchase or a nine-ninety-nine yearly subscription, with one purchase covering iPhone, iPad, and Mac. It needs iOS eighteen, iPadOS eighteen, or macOS fifteen Sequoia or later, and supported inputs cover Canon CR3 and CR2, Nikon NEF, Sony ARW, Fuji RAF, Olympus ORF, Panasonic RW2, and Adobe DNG including iPhone ProRAW.

Milo: Community evidence here is thin, though. The App Store listing has one rating of five out of five, and one reviewer called it a great idea but said the app had issues. The developer responded that a photo deletion bug was involved, so this one's still early in its public track record.

Mia: There's also a tool for a very specific kind of anxiety. On Product Hunt, Team Oranges and Lemons launched WIT, short for Will It Travel, an AI communication check for founders and global teams. You paste an email, Slack message, or product copy, and it shows how that text might land across American, Indian, and Singapore English, highlighting potential misunderstandings, explaining where they might come from, and suggesting clearer wording for an international audience.

Milo: The makers frame it as a language risk audit tool, and they're careful about what it doesn't do. WIT doesn't declare one variety of English correct, doesn't rank one above another, and doesn't automatically replace what you wrote — it provides reasoning and an alternative, and you decide what to keep. The reasoning is personal: a maker named Maria said it came out of her own habit of anxiously rereading messages before sending them in international teams. And the team is honest this is an early version, shipped on purpose to gather real feedback before improving it next.

Mia: The community discussion raised two sharp criticisms. One commenter who sends copy to a dozen countries said they usually only spot a confusing line when someone asks what they meant, and asked which phrase trips people up most. Another praised the decision not to auto-replace writing, but warned that false positives are the real risk — if WIT flags a sentence that was actually fine, people stop reading the flags. That same commenter called the choice of three Englishes curious and said they'd expected British English first, and someone else asked whether the tool works in reverse too.

Mia: The 1752vc Pitch Deck Analyzer is getting a lot of founder attention after its Product Hunt launch. It's built by 1752vc, a California venture firm focused on go-to-market, which says it fields more than four thousand startup applications a year. The maker describes the tool as investor-grade feedback in minutes — it reads an uploaded deck slide by slide and scores it against more than three thousand investor attributes, trained on more than twenty-five thousand real decks and the investment decisions that followed.

Milo: The stage calibration is the interesting part. The analyzer checks narrative cohesion and cross-checks claims for contradictions, and the maker says the most common problem it catches is a deck arguing against itself. The average deck comes back with about a dozen flags, separated by severity and tied to the slide that caused them. The maker says only two percent of the flags raised so far were cosmetic — and that this was deliberate.

Mia: More than two thousand founders already used it in alpha before the public announcement, and the maker says decks that clear the bar can lead to an actual investment from the firm. The launch offer gives five free deck reviews with a promo code and no credit card required. Commenters who tested it reported genuinely useful results — one said a pre-seed analysis surfaced everything from missing technical lead context to unsourced figures, with critical fixes and quick wins clearly separated. Another noted most flags pointed at missing information, but it did account for the startup's stage.

Milo: One open question from the community is whether it stress-tests the full financing logic, and whether it flags decks that are overloaded or under-told. So the resolved feedback is solid, but those deeper checks are still unconfirmed.

Mia: Separately, a browser-based spy satellite simulator called God's Eye View is drawing attention. The maker describes the sources as public and the data as real. It renders a photorealistic 3D globe with live aircraft, ships, satellites, earthquakes, traffic, and public cameras, and adds hands-free voice control through a realtime AI agent. It's fully open source under the MIT license, runs in the browser, and is designed to cost little or nothing for personal exploration.

Milo: The stated point is for people to break it, remix it, and add layers — pull requests are welcome. And the community has been engaged: one commenter said initial demos went viral on YouTube, and the most common viewer request was literally "give us the code." But that same attention raised accuracy questions that remain unresolved.

Mia: Right — one commenter asked directly whether the data is one hundred percent accurate, while another drew a sharp line between live satellite imagery or telemetry and a real dataset that isn't updating in real time, calling those very different demos. So whether the feeds are genuinely live or a static real dataset hasn't been confirmed. A third commenter warned that people may file it as a toy and never check whether an aircraft position is current, and suggested putting the source and timestamp on the globe itself rather than burying it on an about page.

Mia: We covered some interesting releases today, from Tencent's open 770-billion-parameter model built for long-horizon agentic work to Cohere's Parse 5, which turns messy enterprise documents into structured data for AI agents.

Milo: Great to dig into both. Thanks for listening, and we'll catch you next time.