True, but we believe that with the introduction of GGUF to WebGPU, we’ve solved the chicken-egg problem (we tried WebGPU ONNX models before with transformers.js from HG with ~50× fewer model choices). GGUF community is more mature, with a better choice of models quantizations as well.
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).
it was our first post, sorry, we are a small team of ds and developers, and this is our first step outside of closed beta. you can follow us on twitter and telegram for more often updates (links are in the blog).
You're absolutely right (lol, I'm talking like gpt4o now), after initial feedback we found that there is a better way to present webgpu gguf engine without registration — some kinda webgpu gguf playground and we are currently building it!
(if you have any neat ideas — I would love to hear it)
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.
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).