Looks like this is a wrapper around: https://github.com/mlc-ai/web-llm
Which has a full web demo: https://chat.webllm.ai/
Is this correct?
It doesn't seem so to me, either from the way it works or from what little of the code I've looked at...
But I don't have time to do more than the quick glance I just did at a few of the files of each and need to run, so hopefully someone cleverer than me who won't need as much time as me to answer the question could confirm while I'm afk
Entirely correct, c.f. LLM.ts. The bit to Show HN here, is about ~40 lines of code. Simple Typescript MVP of calling the library.
(source: maintains an LLM client that works across MLC/llama.cpp/3P providers; author of sibling comment that misunderstood initially)
Thanks! (Have to admit your sibling comment confused the hell out of me.)
I am to see it regardless - projects been very low activity for months. Just last night I was thinking about ripping it out before launch. No observable future.
EDIT: Ah, wrapper ~= using it as a dependency, no more
Fun demo but the model that's used seems to be pretty stupid:
> What's the best way to get to space?
>> Unfortunately, it's not currently possible for humans to travel to space in the same way that astronauts do. While there have been several manned missions to space, such as those to the International Space Station, the technology and resources required to make interstellar travel feasible are still in the early stages of development.
>> Currently, the fastest spacecraft is the Apollo missions, which took around 77 hours to reach the Moon. From there, they traveled to other planets, but the distances between them are so great that even at high speeds, it would take many years to reach them.
>> As for traveling to space itself, there are no commercial or private companies that offer space travel services, and the cost of such a journey is estimated to be in the tens of millions of dollars per person.
>> However, there are some concepts and technologies being explored that might one day make space travel more accessible and affordable. For example, private companies like SpaceX and Blue Origin are working on developing reusable rockets that could potentially take people to space more quickly and reliably.
>> But for now, let's just say that space travel is still in the future, and we'll have to wait a bit longer for humans to venture into the cosmos!
Are you kidding me? 500MB? And it’s that coherent?
It shows an interface of what could be.
It works on small Llama-3.2-1B model, specially for less powerfull GPU devices
The answer is still terrible for the model size. Maybe it's the 4 bit quantization, smaller models tend to react worse to that
For reference, [1] is what stock quen3-0.6B would answer. Not a perfect answer, but much better at nearly half the number of parameters
1: https://markdownpastebin.com/?id=7ad4ad9f325d4354a858480abdc...
It's likely the quantization on "Llama-3.2-1B-Instruct-q4f16_1-MLC". inference.net generated this more coherent answer: https://hst.sh/ovilewofox.md
It’s broadly correct, if you interpret ‘humans’ as ordinary humans who are not astronauts.
That generous redefinition.of “humans” deals with only the error in the first sentence of the response; the rest is still not “broadly correct” even with it in place.
Blue Origin is built on taking ordinary humans to space. So is/was Virgin Galactic, though they are in a bit of a transitionary phase right now. SpaceX is also willing, they might even take you on a flyby of the moon if you bring money and patience (iirc the last customer lost patience). Basically just ring up your favorite multi-billionaire's space program. And while the estimated price would be correct for SpaceX, rumored prices for Blue Origin's New Shepard are only in the hundreds of thousands per seat
Edit: also the "but if you do that you are an astronaut, so it's still true that only astronauts can do that" loophole was closed when the FAA redefined the word astronaut in 2021. At least if you follow their definition of the word
There is a Prompt API in development that's available in both Chrome and Edge to give access to a local LLM. Chrome extensions have access to it and I believe websites can request access as part of an origin trial.
The model is fully managed by the browser. It's currently the Gemini Nano model on Chrome, and they are testing a version of the Gemma 3n model in beta channels. Edge uses phi-4-mini.
More information is available here: https://github.com/webmachinelearning/prompt-api
This demo only works if you have the webgpu feature "f16". You can find out if you have this by checking for the feature list in https://webgpureport.org/ . The page itself can of course check for this but since f16 support is common they probably just didnt bother.
Also here https://caniuse.com/webgpu
I asked "Why is the sky blue?" and got back a response of
"coppia RR TalentDevExpressincer+'.//////////////////////////////////////////////////////////////////////// cha ولا.AutoSizesaving proleงคicate Like"/>
infos эти za cornerback economical (%]\ enumligne.execRELEASEPropagation_station Bucks проHEME seas GASPOST[Unit(suffix Gloves"
(and so on, for a few more paragraphs).
Am I missing something?
Set llm-auto-lobotomy to disabled in chrome:flags. I think this is only available in canary at the moment
Well, at least it is more informative than an answer like "42".
> Am I missing something?
Realistic expectations.
It seems to be working for some people. I'm just curious whether there was something I could change to make it work.
fuck me it’s hacking the internet
Model used: Llama 3.2 1B (small). Quality hould be similar with running Ollama app with same small model.
What's the performance of a model like vs an OpenAI API? What's the comparable here? Edit: I see it's same models locally that you'd run using Ollama or something else. So basically just constrained by the size of the model, GPU and perf of the machine.
Yes, its very similar to Ollama app, and Llama-3.2-1B model used
Very cool!
For some awesome examples of running different types of models in webgpu check out the examples folder here. https://github.com/huggingface/transformers.js
To have a gpu inference, you need a gpu. I have a demo that runs 8B llama on any computer with 4 gigs of ram
Any computer with a display has a GPU.
Sure, but integrated graphics usually lacks vram for LLM inference.
Which means that inference would be approximately the same speed (but compute offloaded) as the suggested CPU inference engine.
Does anyone know why all of these WebGPU LLM demos have you download the models to browser storage rather than letting you open a gguf already on your local drive? I have several models downloaded already that I would be interested in trying.
Browsers are sandboxed away from user storage.
You can change this by changing settings, command line arguments, build flags, etc. But can’t really expect people to do this just to use your website.
You can open a file for performant access in all major browsers. It's the same API used for uploading files (<input type="file" />), but you can also just load them into memory and do stuff.
Sure, with the caveat that the file is specifically selected by the user.
Which sounds ideal for this scenario, right?
I read that as the ask, yes.
WebGPU's security model restricts direct file system access, requiring models to be loaded through fetch/cache APIs rather than local file paths.
I did this with mlc @ https://wiz.chat some time ago.
Warning: it has a llama 3.1 7b model and is around 4 gb. It needs either a GPU or a Macand works only on chrome
Its not working.
Loading model: waiting WebGPU error: WebGPU is NOT supported on this browser.
Im on macOS and Safari.
Same on iOS and mobile Safari - not working.
time to vote with your wallet, Safari does not support WebGPU.
I've been following the development of WebGPU and its potential applications, and this demo is a great example of what's possible.
Any way to make this use my actual GPU instead of the terrible integrated one?
Which model does the demo use?
Llama 3.2 1B
Cool idea but badly broken from a little testing.
very cool. improvement would be if the input text box is always on screen, rather than having to manually scroll down as the screen fills.
Beware of opening this on mobile Internet.
Demo site is asking before download
Well, I am on a mobile right now, can someone maybe share anything about the performance?
Not everyone enjoys unlimited data plans, and 500 MB is schon a lot.
On my Android device works pretty fast.
But keep in mind that it's small Llama-3.2-1B model, specially for less powerfull GPU.
Now there is an error