TheaterFire

7B UI Model that does charts and interactive elements

Posted by United-Rush4073@reddit | LocalLLaMA | View on Reddit | 34 comments

7B UI Model that does charts and interactive elements
[https://huggingface.co/Tesslate/UIGEN-T2-7B-Q8\_0-GGUF](https://huggingface.co/Tesslate/UIGEN-T2-7B-Q8_0-GGUF)

Reply to Post

34 Comments

Thicc_Pug@reddit

ok hear me out, what if instead of one huge model, we train many smaller domain specific models like this. Then we submit the promt to a "master" model which decide which domain model to use. The "communication" between domain models and "master" model doesnt even have to be words, but raw tensors. In the end product, the user can decide which domains they want to be able to use.
View on Reddit #55100991

ExplorerWhole5697@reddit

This is a great idea. Each model is an expert at some tasks. The whole thing becomes a mixture of experts. This could potentially lower RAM requirements because not all experts need to be loaded in memory at the same time.
View on Reddit #55181213

myvirtualrealitymask@reddit

Do you plan on doing something similar with the qwen3 models?
View on Reddit #55083318

United-Rush4073@reddit (OP)

Absolutely, thats next on the list!
View on Reddit #55083354

AaronFeng47@reddit

I saw some people use LLMs to generate webpages, then take screenshots of them to use as a PowerPoint presentation. Maybe you can train a "PowerPoint" model like this? Since you guys are really good at training LLMs to do UI design.
View on Reddit #55084645

United-Rush4073@reddit (OP)

This is really interesting, I'm thinking about it right now. Because I do know that you can use python to make pptx pages as well, and recently there was that awesome icon SVG model. Maybe all 3 can be combined in a workflow to make the better powerpoints? I'm brainstorming.
View on Reddit #55087112

kmeansneuralnetwork@reddit

This would be really helpful for people like me who hate presentations.  Also, I am completely new to this domain, so, i am sorry if i am wrong. You mentioned RL with help of UX researcher. Doesn't people normally do RL with bigger models and evaluate it? 
View on Reddit #55150343

Vivid_Dot_6405@reddit

That'd be awesome!
View on Reddit #55085347

Ylsid@reddit

Oh, this is cool. I'd love something like this for other Frameworks too, especially those with irritating UI setups
View on Reddit #55149941

United-Rush4073@reddit (OP)

The latest version of UIGEN-T2. This is a culmination of everything we've learned since we started, pulling together our reasoning and UI generation. We have a new format for reasoning, that thinks through UI principles. Our reasoning was generated using a separate model and then transferred. More details are on the model card and the link to it. We've also released our [LoRas](https://huggingface.co/Tesslate/UIGEN-T2-7B-LoRAat) each checkpoint, so you don't have to download the entire model, as well as make your own decision about which version you like. You can download the model here: [GGUF Link](https://huggingface.co/Tesslate/UIGEN-T2-7B-Q8_0-GGUF) In the near future, we plan on using this model as a base for reinforcement learning, but we are looking for resources to do that. If you want to demo without downloading anything: [Playground](https://huggingface.co/spaces/smirki/UIGEN-T2-Playground), [HF Space](https://huggingface.co/spaces/smirki/UIGEN-T2-7B-Artifacts-Demo) And we didn't find any good (simple) Artifacts demos, so we released one in Open Source: [Artifacts](https://github.com/TesslateAI/UIGEN-T2-Artifacts)
View on Reddit #55081803

lochyw@reddit

using up to date tw4? Most of the code generators use super old code references that you wouldnt bring into a project. Also typically would build them via templates in combination with some backend lang like go or something, so curious about getting something actually usable vs just a nice design preview tool
View on Reddit #55149217

vulture916@reddit

Both demos give GPU errors.
View on Reddit #55082245

United-Rush4073@reddit (OP)

Thanks for letting me know! I did feel like it was a little risky but it can only go as far as the ZERO gpu timeout lets it. I'll apply for the HF Spaces prgram.
View on Reddit #55082458

kyleboddy@reddit

Running into the GPU errors as well. > The requested GPU duration (600s) is larger than the maximum allowed
View on Reddit #55099017

United-Rush4073@reddit (OP)

Seems to be giving me an output on desktop so maybe its broken on mobile. https://preview.redd.it/ravc1uy9syxe1.png?width=1537&format=png&auto=webp&s=de239953ff97989658e31f23ab25af225da6f9fe
View on Reddit #55082678

nic_key@reddit

Any chance that there will be quantized GGUF as well (potentially uploaded to Ollama?)
View on Reddit #55083957

United-Rush4073@reddit (OP)

You can use this dropdown on huggingface to bring it into Ollama! [https://huggingface.co/Tesslate/UIGEN-T2-7B-Q8\_0-GGUF](https://huggingface.co/Tesslate/UIGEN-T2-7B-Q8_0-GGUF) https://preview.redd.it/qcksyzlzvyxe1.png?width=1612&format=png&auto=webp&s=88c1695ed2bf0d11c76f93612e53398a821a073d
View on Reddit #55084101

nic_key@reddit

Thanks!
View on Reddit #55084146

fnordonk@reddit

I don't do front end work your project has been inspirational in what can be done with loras. Happy to read your team is growing.
View on Reddit #55083840

United-Rush4073@reddit (OP)

Thank you! Yeah, we're working on all kinds of cool things, stay tuned!
View on Reddit #55086971

drfritz2@reddit

Can you provide links to your work and methodology? I want to make a model that has knowledge about an online game. For E-sports And another project for education, but I know almost nothing about how this is actually done
View on Reddit #55138972

Boring_Resolutio@reddit

how can we follow your journey?
View on Reddit #55100768

v1sual3rr0r@reddit

What is the license?
View on Reddit #55137856

verticalfuzz@reddit

Is it possible to run this in openwebui/ollama?
View on Reddit #55137734

Muted-Celebration-47@reddit

I found UIGEN T1.5 has 32B. Do you plan 32B version for UIGEN T2?
View on Reddit #55128997

seeKAYx@reddit

This is exactly what I was looking for. I still use Sonnet 3.5 for front end. But this could replace it. Thank you m!
View on Reddit #55124568

davidpfarrell@reddit

Looking forward to trying this out. BTW: DevQuasar has quanted this out to various sizes based on the BF16 model: [https://huggingface.co/DevQuasar/Tesslate.UIGEN-T2-7B-GGUF](https://huggingface.co/DevQuasar/Tesslate.UIGEN-T2-7B-GGUF) Thanks for sharing!
View on Reddit #55113608

Danmoreng@reddit

Wondered when anyone would do something like this. Of course there is v0 and vue0 for react and vuejs components, but they use openAI without specialised models as far as I know. Do you plan to do similar training for other frameworks? I read bootstrap, which is a nice start. I kind of dislike tailwind because of the bloated HTML too many style classes produce. Would love to have some specialist model for vuejs components, probably even without a CSS framework.
View on Reddit #55100806

JLeonsarmiento@reddit

Excellent.
View on Reddit #55084128

Willing_Alarm_4192@reddit

Agreed
View on Reddit #55098265

Big-Helicopter-9356@reddit

This is phenomenal. How did you design your RL rewards?
View on Reddit #55094742

Anarchaotic@reddit

This looks great! I've had trouble with Pytorch before since I'm on a 5XXX series GPU (Blackwell) - do you happen to know if this will work with a 5080/5090?
View on Reddit #55092187

FullstackSensei@reddit

Really like this! I think this is the future of small LLMs. Any chance you'd release your training pipeline and dataset, similar to what [oxen.ai did with Qwen 2.5 Coder 1.5B](https://www.oxen.ai/blog/training-a-rust-1-5b-coder-lm-with-reinforcement-learning-grpo) and [Together.ai with DeepCoder](https://www.together.ai/blog/deepcoder)? I see you also have Rust fine tune (Tessa) and have released the datasets for that. Any write ups on Tessa? Any chance you'd release the training pipeline? Would be very interesting to see how well it would replicate with a 1.5B class model.
View on Reddit #55090653

ThiccStorms@reddit

great, i can finally not worry about frontend dev.
View on Reddit #55085272