Gr1m

Crypto and tech enthusiast. BTC maximalist.

[Gr1m @GrIlm14]

Finally got my hands on 2x @nvidia dgx sparks. I’ve got them set up and running Deepseek v4 flash NVFP4. Any other model suggestions?

[Gr1m @GrIlm14]

@ollama @MichaelGannotti Gotta be pi

[Jun Song @jun_song]

Great move as I expected days ago!

Nvidia will lead US open-weight AI soon.

[Gr1m @GrIlm14]

Getting 2 sparks has been such a game changer for me. Local ai is the wave rn! 🔥

[Gr1m @GrIlm14]

@jun_song Been waiting to see something like this!

[0xSero @0xSero]

Weekend project to level you up.

Cost: 50$~
Time: 4-8 hours
Goal: Train GPT2 level model
Language: Python

You'll learn:

1. Datasets
2. Chinchilla
3. Provisioning Compute
4. Training
5. SFT
6. Benchmarks

Karpathy's YouTube series is excellent alongside
https://t.co/Ts4iGGYR6Z

[Gr1m @GrIlm14]

@WescheNex1q Great comparison

[Daegwon Nacyot Kim @nacyotKim]

"DGX Spark에서 LLM 추론 중 가장 뜨거운 건 GPU가 아니라 CPU였다. py-spy로 추적하니 vLLM의 대기 루프가 코어 4개를 최대 클럭으로 헛돌리고 있었다. 기본값 한 줄을 고쳐 CPU 낭비 12분의 1, SoC 최고 97→73°C. 성능 손실��� 없다."
https://t.co/fDEhfrqAEs

[Gr1m @GrIlm14]

@MiaAI_lab Oooh! This is exciting. Can’t wait to see what you’ve got cooking here

[Gr1m @GrIlm14]

@dotsstudioai 👀

[Jun Song @jun_song]

Qwen3.8-max is open now.

Just a heads up, ModelScope is run by Alibaba, same as Qwen.

Looks like the 27B model isn't out yet, but I heard it's coming this Friday.

Open-sourcing a massive 2.4T model like this is no easy task.

[TechMD @TechMDAI]

If you have a spark, This is it https://t.co/U6u1ijbmml

[Mia @MiaAI_lab]

sparkDash ⚡️ now watches ComfyUI

v1.6 is out - monitor your comfyui jobs with style.

• model/loras footprint, steps, nodes, and more
• live job queue + model footprint (res · steps · sampler)
• progress bar
• last finished job + duration
• cancel / remove from the dashboard
• open comfyui in one click
• overview chip: idle · run · Nq
• queue eta
• ui improvements & fixes

→ https://t.co/8H3h5wjPOh

[Gr1m @GrIlm14]

@shannholmberg I need to try this

[Gr1m @GrIlm14]

@Dave_Charland @MiaAI_lab I did the same! I ordered 2 from Amazon (micro center was the seller) and both were brand new/have worked well so far. Micro center all the way!

[Gr1m @GrIlm14]

@TrickRiggin @mr_r0b0t Would love to copy this

[Killy @net_termina]

For NVIDIA/PNY DGX Spark owners out there, could you share your experience on how you deal with high temperature? Is overheating a common issue to begin with, and if so, how does it impact performance? I’m looking to purchase one for minimum viable build demo and was wondering paying extra for Gigabyte model that apparently has better thermals is worth it. In fact, if anyone has a Gigabyte model, do you mind sharing if that’s true?

[0xSero @0xSero]

Yours to keep.

https://t.co/aJY3UV2SA8
https://t.co/YTaS1SfpxH
https://t.co/vH2RdS2NQT

[Jun Song @jun_song]

First impression of testing SuperDeepseek-V4-Flash-0731 on 2xDGX Spark :

I might not need to subscribe any AI anymore.

Just amazing

[Gr1m @GrIlm14]

@4RConsultancy @mr_r0b0t This set up is sick! What models are you using? I’m specifically curious about the openclaw and Hermes performance

[BridgeMind @bridgemindai]

Grok 4.5 is the most underrated model in AI right now.

Speed, cost, and intelligence.

Nobody else balances all three.

It scores near the frontier at $1.51 per task, finishes in half the steps of the competition, and I ran it all day on SuperGrok and used 1% of my weekly limit.

Every other lab makes you pick two.

And Grok 4.6 is coming. If they hold this balance and push the intelligence, SpaceXAI has a real shot at the top.

Very excited to test it.