MDST Engine: run GGUF models in the browser with WebGPU/WASM
mdst.app
mdst.app
Just in case (a common question): we are planning to open-source the MDST WebGPU engine, but please give us some time to polish it and get it ready for a harsh oss community.
Meanwhile, you can skip the waiting list with this invite: hackernews_Eq1RDox
WebGPU technology gives us a way to provide truly free and accessible user tiers for everyone without burning investor capital or doing rug pulls.
If you have any questions, I’d love to answer them.
I don't think there's a whole lot of rush on your end, given the level of support WebGPU generally has at this point.
Also, with better consumer hardware, smarter small models (up to 4B), and better WebGPU engines and competition between them (MDST Engine is still in a very early stage, with little optimization, no flash-attention yet, etc.), we think this technology will grow super fast in 2026 for a less tech-savvy market (which we believe is good for open-weight models).
MDST is exciting and so was this blog post. I'd like to read more.
Also, since we are forced to sign in, there is no option to use the website for auth and only Google or Github, both not the best choices.
(if you have any neat ideas — I would love to hear it)
In this synthesized machine states era do I want to still run a bloated, buggy web browser?
I think I will keep working on synthesizing a replacement for the desktop metaphor of a blank viewport/canvas and sampling from a model of vision data, tweaking at the edge when labels are missing or wrong, edges don't align, saving it back to model
Think "boot to Blender" or Unreal Engine, synthesizing desktop and text files and such inside that environment
There is increasingly little reason to import all the state of legacy software beyond a minimal Linux distribution
Prompt to binary is on the way, skipping the generation of the intermediary code layer altogether
Code is legacy technology from the 1950s era of using machines; arbitrary syntax for labeling with human context the use case of a memory address
No reason the labeling has to be at the machine level. That can be vectors and labels of human value applied right to what's shown onscreen
What tech people should be focused on is political action that keeps hardware open and models open and not locked behind data center.
Software dev as we knew it is already over technology wise, social acceptance has yet to propagate, but social acceptance is an eventually consistent system. It’ll happen
I do. The browser is indispensable, at least for now, and it's much better to keep improving it for the age of AI than to jump to fairy tails and unproven replacement technologies.
> What tech people should be focused on is political action that keeps hardware open and models open and not locked behind data center.
And while they're working on this, you're going to do what exactly? You made a lot of forward looking claims that have nothing to do with the present reality, you just forgot to present any realistic vision about how to get there.
If you can't imagine procuring any those pink unicorns yourself why do you plead with others to deliver them to you? It just doesn't add up.
I mentioned I am working on replacements for the usual local PC software stack ..."exactly"
You breezed right over that, obsessed over a single statement in a rush to justify your choices.
In my day to day experience the browser is not indispensable, quite the opposite. I barely use one. Native phone apps work better than web apps for "present reality" use-cases. I will stick to phone apps and run what I want on my PC.
I am waiting for a new build of a vision model backed GPU accelerated Vulkan-based desktop replacement to compile.
I can submit a prompt, get a point cloud/depth map displayed onscreen and then it crashes. Memory management needs improvement.
And I have achieved this in just a few weeks with a single 3090 and local models. But I guess being an EE who has developed software from hardware up for over 20 years and not a web SaaS code school grad, has provided the context to bring "fairy tales" to life.
Meanwhile Web Assembly has been right around the corner for over a decade; a fairy tale with little value in present reality.
Very good, when you replace it, use it to create all the shiny things you dream about, it should be a breeze, use AI to create everything you need. We, the less fortunate, are going to stay with the browser only until you save the world and not a minute longer.
> I will stick to phone apps and run what I want on my PC.
I also run what I want on my PC and what I want in my browser and I don't want phone apps because they are restricted to a tightly controlled environment. Are you going to replace that one too? With something free and open? AI should do it with a couple of prompts, right?
> And I have achieved this in just a few weeks
That's wonderful, I wish you luck, seriously.
> EE who has developed software from hardware up for over 20 years... has provided the context to bring "fairy tales" to life.
Until you bring them to life, they remain fairy tales - not my call, that's how it works and has always worked. I don't see why we should discuss something that isn't there and address claims that cannot be verified in any way.
More importantly, I don't see why we should abandon vital and indispensable infrastructure just because you say so and because you can get by using something much more restricted for your narrow purposes. I mean, show me, don't tell me.