Mia

Local AI, LLMs, tech thinker & builder

[Mia @MiaAI_lab]

I just published Slate — a fast, light-weight OLED-friendly Markdown/text editor.

It supports editing all types of text-based files.

One thing I really wanted: a proper OLED-friendly editor. Not “dark gray” — complete black, so it looks great on OLED displays and feels easy on the eyes at night.

Fully developed by local AI.

Currently Windows only. Feel free to fork and build for Mac/Linux.

Feel free to test it, open issues, report bugs, or suggest ideas.

https://t.co/KImPGEHmvp

[Mia @MiaAI_lab]

@advented_ Do you mean single session? Unfortunately I'm not getting 40-50 t/s, it was unstable. My recipe gets me to 31.7 t/s for single session.

For multiple sessions I get up to 80 t/s

[Mia @MiaAI_lab]

53 tok/s achieved on Step-3.7-Flash NVFP4 with MTP on 2x DGX Spark with 256k context. 🎉🥳

Elapsed time: 56.694 s
Prompt tokens: 29
Generated tokens: 3000
Total tokens: 3029
Generation tok/sec: 52.92
End-to-end tok/sec: 53.43

[Mia @MiaAI_lab]

@ArnauBosch10 Thanks appreciated :)

[Mia @MiaAI_lab]

Final results for are in. 53k tok/s is NOT stable. Had to do some tunings to make it work.

Eventually the best results are around 33 tok/s.

This is Step-3.7-Flash nvfp4 MTP & no --enforce-eager.

Repo will be published in a few hours.

[Mia @MiaAI_lab]

53 tok/s achieved on Step-3.7-Flash NVFP4 with MTP on 2x DGX Spark with 256k context. 🎉🥳

Elapsed time: 56.694 s
Prompt tokens: 29
Generated tokens: 3000
Total tokens: 3029
Generation tok/sec: 52.92
End-to-end tok/sec: 53.43

[Mia @MiaAI_lab]

@TeksEdge @MichaelGannotti @NVIDIAAI Where?

[Mia @MiaAI_lab]

@barackomaba @ChicouTiMix @NVIDIAAI @0xSero nvfp4, 40-45 tok/s

[Mia @MiaAI_lab]

Codex app on Windows running DeepSeek-v4-Flash through Codex Shim, running on 2x @NVIDIAAI DGX Sparks. @0xSero

Works so well...

[Mia @MiaAI_lab]

@mr_r0b0t Yeah multiple sessions are really good

[Mia @MiaAI_lab]

Finally cleaned up and published my Dual DGX Spark setup for Step-3.7-Flash-NVFP4.

The repo includes scripts for 2x DGX Sparks, vLLM, no-MTP first, optional MTP grafting, live logs during startup, background serving after ready, plus stop/status/test helpers.

~33 tok/s, even when in deep context.

Hopefully saves someone else a few hours of pain 🙂

Repo: https://t.co/10QcqYKK9Q

[Mia @MiaAI_lab]

@ChicouTiMix @NVIDIAAI @0xSero Yes it's better

[Mia @MiaAI_lab]

@ainslec @AMD @NVIDIAAI 4 x yes

[Mia @MiaAI_lab]

@naturalSmartnes Yes the only reason they are doing it is to lure Claude users to Codex. That's it.

[Mia @MiaAI_lab]

For how long OpenAI can keep resetting limits? It's obvious they are doing it to get users from Claude to Codex.

But I don't think it's sustainable.

[OpenAI @OpenAI]

We heard you wanted to use Codex rate limit resets on your own time.

Starting today, we’re rolling out the ability to save rate limit resets to use later.

We’re starting Go, Plus, Pro, and Business users with one free reset:

[Mia @MiaAI_lab]

@ainslec @AMD @NVIDIAAI It's not just better. I can connect up to 8x DGX Sparks and upto 1tb of unified memory.

[NVIDIA AI @NVIDIAAI]

Congrats to the @MiniMax_AI team on the release of MiniMax M3, a long-context multimodal model for text, image, and video reasoning. 🙌

Try it today with our free GPU-accelerated endpoint on https://t.co/es07MrU5I0.

Details: https://t.co/89qlcTP3OW

[Mia @MiaAI_lab]

@0xSero I'd like to know how it compares to DS4-Flash. Awesome job.

How can I get access to the site?

[Mia @MiaAI_lab]

@emilstridell @NVIDIAAI Haha I think it was a spike, but in general DGX Spark has insane prefill speeds, literally.

[Mia @MiaAI_lab]

Two concurrent sessions with DS4-Flash, getting more than 60 tok/s and insane prefill numbers.

Running on 2x @NVIDIAAI DGX Sparks

[Mia @MiaAI_lab]

@nivkorin2004 @NVIDIAAI Yes

[Mia @MiaAI_lab]

@spark_arena This is awesome! Will try this tonight.

[Mia @MiaAI_lab]

Building the things you couldn't find anyone else building has never been easier.

[Mia @MiaAI_lab]

👀

[Mia @MiaAI_lab]

@mikhei777 For one DGX Spark I'd go with Qwen 3.6 35b NVFP4 with MTP, with full context, or DS4-Flash-REAP from @0xSero

[Mia @MiaAI_lab]

For the price of 2x DGX Sparks you can run frontier open models with full context in very decent and usable speeds for agentic coding.

30-45 tok/s on frontier models, NVFP4 + MTP.

No usage limits.

I don't think there is any better deal out there for 256gb available vram.

[Mia @MiaAI_lab]

@ArnauBosch10 Thanks 🙏🏻