Cursor, but for COBOL

gm legends, happy Tuesday.

A YC startup built Cursor for mainframe developers; OpenBMB fit a vision model into 1.3 billion parameters that runs on a phone; Kelviq wants to be the one button that handles payments, tax, and billing for SaaS teams; and we're announcing the Vercel Day winners and opening the door for Friday's.

Hopper is built by Hypercubic (YC F25), founded by a former Apple ML engineer who shipped multi-modal algorithms to 200M+ devices and decided the most important unsolved problem in software was COBOL — it's an agentic dev environment for mainframe systems that can inspect data, edit code, write JCL scripts, submit jobs, and diagnose failures autonomously.

🔥 Our Take: There are still billions of lines of COBOL running quietly inside every major bank and insurance company, maintained by developers using 1970s-era interfaces while everyone else got Copilot and Cursor. The founder story here matters: this isn't a developer who stumbled into mainframes, it's someone who shipped Apple-scale ML work and chose this as the problem worth solving. That's a different kind of conviction.

The results are in from our first Vercel Day. QA Tech took the top spot and will pitch directly to Vercel Ventures. Five runners up — Stoa, CalendarPipe, Athena, Aria, and Arky — each walk away with $30k in Vercel perks.

This Friday, May 15, is the next one. Launch your product on Product Hunt and you're on the official Vercel Day leaderboard, in front of thousands of builders already paying attention. Top launches win prizes.

Schedule your launch now. Three days left.

MiniCPM-V 4.6 comes from OpenBMB, the research lab out of Tsinghua University's NLP group that MIT Technology Review named one of the Chinese AI teams to watch after DeepSeek — and this release is a 1.3 billion parameter vision-language model built to run entirely on-device on mobile hardware, no cloud required.

🔥 Our Take: Most people think of vision models as cloud-only, GPU-heavy things. This one fits on a phone. 1.3B parameters doing image understanding that normally takes models 100x its size. The team has been publishing on model compression for years, so this isn't a one-off — it's a research direction. If it works as described, the privacy case writes itself.

Kelviq is built by Sachin Neravath, who spent five years running ParityDeals and had hundreds of conversations with SaaS founders about billing pain — Kelviq is what he built from all of that: a Merchant of Record platform that handles payments, global tax, usage-based billing, license keys, and compliance in one place at 3.5% + 40¢ per transaction.

🔥 Our Take: The Merchant of Record model is the key detail — Kelviq takes on the tax liability, not just the payment processing. That's a different product from Stripe. For AI companies with usage-based billing across multiple countries, the compliance overhead is genuinely the nightmare part. Five years running a SaaS pricing tool means Sachin has seen this problem up close more than most people building payments infrastructure.