Wouldn't it be nice if we have language specific llms that work on average computers.

Like LLM that only trained on Python 3+, certain frameworks, certain code repos. Then you can use a different model for searching the internet to implement different things to cut down on costs.

Maybe I have no idea what I'm talking about lol

I imagine some sort of distill like this would be possible, but I think multi-language training really helps the LLM.

How good is ethernet over electrical sockets these days. I had one about 15 years ago maybe, but it wasn't that good.

Has tech changed. I'd use it over my wifi setup if its was fast.

This tech is known as BPL(broadband over powerline) if you want to look further into it.

I used the regular home power stuff many years ago and the speeds were pretty bad and the network loss was unreliable.

My understanding is that it has improved in some circumstances, but if the connection ends up "hopping" through your breaker you get back to garbage speeds.

In theory you can get 2 Gbps speeds, but in practice it seems like still around 500 Mbps. I don't know if the loss has improved but it was a significant problem before, since even a low loss will render a connection unusable.

Its easier to adopt AI when starting from scratch or code base is well maintained.

This is unreal. Nice work How long did it take ? I tried to use ash to build a simple app and couldn’t get it to work lol.

I’m an elixir noob

A few weeks (mostly weekends)

How are you using the huge models locally?

First time I tried it, claude built all the files in the wrong directory lol. It's working fine now.

What happens if I want to make the video on the fly and save that to reuse it when the same question or topic comes up. No need to render a video. Just play the existing one.

This isn't natively supported -- we are continuously streaming frames throughout the conversation session that are generated in real-time. If you were building your own conversational AI pipeline (e.g. using our LiveKit integration), I suppose it would be possible to route things like this with your own logic. But it would probably include jump cuts and not look as good.

How did you do this locally? Tools? Language?

I just followed the Quickstart[1] in the GitHub repo, refreshingly straight forward. Using the pip package worked fine, as did installing the editable version using the git repository. Just install the CUDA version of PyTorch[2] first.

The HF demo is very similar to the GitHub demo, so easy to try out.

  pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
  pip install qwen3-tts
  qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-Base --no-flash-attn --ip 127.0.0.1 --port 8000

That's for CUDA 12.8, change PyTorch install accordingly.

Skipped FlashAttention since I'm on Windows and I haven't gotten FlashAttention 2 to work there yet (I found some precompiled FA3 files[3] but Qwen3-TTS isn't FA3 compatible yet).

[1]: https://github.com/QwenLM/Qwen3-TTS?tab=readme-ov-file#quick...

[2]: https://pytorch.org/get-started/locally/

[3]: https://windreamer.github.io/flash-attention3-wheels/

https://github.com/sdbds/flash-attention-for-windows/release... - FA2 binaries for you

It flat didn't work for me on mps. CUDA only until someone patches it.

Demo ran fine, if very slowly, with CPU-only using "--device cpu" for me. It defaults to CUDA though.

Try using mps I guess, I saw multiple references to code checking if device is not mps, so seems like it should be supported. If not, CPU.

Kind of a noob, how would I implement this locally? How do I pass it audio to process. I'm assuming its in the API spec?

Scroll down on the Huggingface page, there are code examples and also a link to github: https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base

I wanted to try this locally as well so I have asked AI to write CLI for me: https://github.com/daliusd/qtts

There are some samples. If you have GPU you might want to fork and improve this, but otherwise slow, but usable on CPU as well.

Dollar milkshake theory