February 14, 2026
Local AI, global drama
MDST Engine: run GGUF models in the browser with WebGPU/WASM
Hype for local AI meets a login wall, and the comments erupt
TLDR: MDST's engine runs AI models locally in your browser, promising privacy and no cloud. Commenters love the idea but clash over a required login and Google/Github-only sign-in, debating 'open-source soon' versus trust today—why a 'local' tool needs an account is the lightning rod.
MDST dropped a flashy promise: run AI models right in your browser with no cloud, using your own computer’s graphics power. The team says their Web tech lets anyone load GGUF files (a popular model format) in Chrome, Safari, or Edge, and boasts end‑to‑end encryption for privacy. Cue the Hacker News crowd assembling like it’s the season finale.
Right out of the gate, a team member pops in with the classic “we’ll open-source it soon” tease and an invite code for skip-the-line access. Some readers are thrilled—finally, local AI that “just works.” But the vibe flips when others point out you must create an account, and it’s Google or GitHub only for sign-in. The real question bubbling up: can you sell “local and private” while asking users to hand over identity just to try it?
Cue the memes: “Local means local—except your login,” “No cloud, just OAuth,” and, of course, “Open-source Soon™.” Fans argue that Web tech keeps costs down so free tiers can exist; skeptics say trust isn’t a roadmap item and want the engine public now, with anonymous access. It’s a classic internet split: one side cheering the tech, the other side side‑eyeing the gate. Either way, the drama is running at full GPU speed.
Key Points
- •MDST released a WASM/WebGPU engine to run GGUF-format models entirely in the browser for local inference.
- •The platform supports Chrome, Safari, and Edge, with Firefox support planned.
- •GGUF is highlighted as a popular, single-file, quantized LLM format suited to consumer hardware and easy tuning.
- •MDST is a free, secure, collaborative IDE integrating cloud and local agentic inference, GitHub sync, and E2E encryption.
- •A Research module offers local benchmarking across tasks and a public WebGPU leaderboard ranking models by weighted scores.