Mia (@MiaAI_lab)

Pinned post

Companies that offer integrated solutions for local inference will boom in the next few years. I'm not talking about another GPU in a case.

Hardware + OS + an inference stack that makes open models easy to run. A plug-and-play system.

Ease of use is still underrated.

Mia @MiaAI_lab 2h

New in my T3 Code fork: adding a project, simplified.

- Click the new 📁+ icon next to New thread
- Pick a machine: this computer, or any SSH host from your config (it connects for you)
- Click through to a folder and press Add project
- Clone or start empty from the same dialog

Mia Verified account @MiaAI_lab 15h

I made some changes to T3 Code to make it local-first.

A new page in Settings > Inference, allows you to directly add your local models, and then attach it to Pi (or others), so it's not depended on what models are configured in Pi on your host.

I've also added an actual Dark theme, and made the previous dark theme as OLED. "Dark" theme was too black for me.

I'll continue modifying it and adding more features.

Full list of changes:
- Local models are the default, cloud is the backup
- Register your own GPU servers, live health & cache
- Link Pi / OpenCode / Codex to a server in one click
- A gateway queues requests, so threads never swamp the GPU
- Automatic fallback when a server goes down
- Titles, commits & PR text written by your local model
- Per-server stats: tokens, cache hits, TTFT, tok/s
- Local runs show "Local", never a fake $ cost
- Lean tool set: ~40% fewer prompt tokens for local models
- Two-tone Dark theme + true-black OLED

Get my fork here:
github.com/MiaAI-Lab/t3co…

View quoted post on X

Mia @MiaAI_lab 7h

This is the best image/video gen suite you can run on your DGX Spark, bare none. Highly recommend you to try it out.

It completely replaced the use of ChatGPT to generate images. It works extremely good on just one spark.

I'll be contributing to it as much as can as I think it's really that good. Kudos to @u1tra_instinct and @SamJWasserman for building this.

github.com/drowzeys/Imagi…

github.com GitHub - drowzeys/Imagination-Studio-by-keys-Wasserman: Imagination Studio by keys+Wasserman... Imagination Studio by keys+Wasserman powered by TensorFold: an AI film studio for Mac and NVIDIA DGX Spark. Plan a film like a crew (Slate workflow by Sam Wasserman), render Qwen-Image-2.1 / FLUX3-...

Mia @MiaAI_lab 8h

Now it's RTX PRO 6000s turn ✨

The new GLM 5.3 Flash update gives you:
- More KV: 973K to 1.05M tokens.
- Survives restarts: conversations are saved to disk and come back in about 1 second.
- Shorter thinking-off answers: no more "Max effort" line when thinking is off.
- Safer startup: a wrong network setting gets fixed automatically.

Update your repo!

github.com/MiaAI-Lab/GLM-…

github.com GitHub - MiaAI-Lab/GLM-5.3-Flash-2x-RTX-PRO-6000-TensorFold: GLM-5.3-Flash EXL3 on 2x RTX PRO 6000 GLM-5.3-Flash EXL3 on 2x RTX PRO 6000. Contribute to MiaAI-Lab/GLM-5.3-Flash-2x-RTX-PRO-6000-TensorFold development by creating an account on GitHub.

Mia Verified account @MiaAI_lab 15h

Another update to GLM 5.3 Flash for 2 DGX Sparks ⚡️

- KV sharding: when enabled, the two dgx sparks split context memory, so you get about 50% more room. This gives you around 3.1M total KV cache.

- Chats now save to disk and resume after a restart. A 101K token chat came back with 98K tokens cached, and the same answer.

- You can now set a minimum prompt cache. If it can’t be met, both nodes refuse to start. I also removed a false warning on every start.

- With thinking off, prompt overhead dropped from 15 to 8 tokens, and garbled repeated tool-call lines are fixed.

Thanks for the community PRs and issue reports that made all this happen at lightning speed!

Get it here:
github.com/MiaAI-Lab/GLM-…
View quoted post on X

Mia @MiaAI_lab 15h

I made some changes to T3 Code to make it local-first.

A new page in Settings > Inference, allows you to directly add your local models, and then attach it to Pi (or others), so it's not depended on what models are configured in Pi on your host.

I've also added an actual Dark theme, and made the previous dark theme as OLED. "Dark" theme was too black for me.

I'll continue modifying it and adding more features.

Full list of changes:
- Local models are the default, cloud is the backup
- Register your own GPU servers, live health & cache
- Link Pi / OpenCode / Codex to a server in one click
- A gateway queues requests, so threads never swamp the GPU
- Automatic fallback when a server goes down
- Titles, commits & PR text written by your local model
- Per-server stats: tokens, cache hits, TTFT, tok/s
- Local runs show "Local", never a fake $ cost
- Lean tool set: ~40% fewer prompt tokens for local models
- Two-tone Dark theme + true-black OLED

Get my fork here:
github.com/MiaAI-Lab/t3co…

Mia

Latest posts 5