Viktor

Love Vibe Coding, MLX, and all the other AI stuff

[Viktor @viktork1289]

@ivanfioravanti Finished after just under 11 hours. The sample images look very promising so far.

I’ll test tomorrow

[Viktor @viktork1289]

Training a Krea 2 LoRA on the DGX Spark.
36 images, 3,000 steps, took about 10 hours.

Curious to see how it turns out

[Viktor @viktork1289]

@ivanfioravanti Should be done in about 3 Hours

[Viktor @viktork1289]

2,000 of 3,000 steps are now complete on my Krea 2 LoRA training run on the DGX Spark.

I’m training a LoRA for a person’s face. About 3 hours left to go, and the temperatures are running pretty high. 🔥

So far, the sample images are looking promising.

[Viktor @viktork1289]

@MiaAI_lab My gx10 Powers automatically on when I switch on my power plug

[Viktor @viktork1289]

@KI_Vater @StefanMaier @NVIDIAAI Ich hatte das Glück vor zwei Wochen einen Asus gx10 mit 2tb für 3700€ zu bekommen

[Viktor @viktork1289]

@KI_Vater @StefanMaier @NVIDIAAI Find jetzt Laguna nicht verkehrt, auch in deutscher Sprache nicht weil ja viele sagen das Model ist hauptsächlich auf Englisch ausgelegt. Hätt auch gern nen zweiten spark aber bei den Preisen zur Zeit ist es einfach nicht drin

[Viktor @viktork1289]

@StefanMaier @KI_Vater @NVIDIAAI Es läuft in 2bit schon auf einem spark aber nur vernünftig mit nem ssd Cache und ich glaub da schreibt man sich die ssd nach ner Zeit kaputt

[Viktor @viktork1289]

@deepseek_ai A flash light version with 120b would be nice for single spark setups

[Viktor @viktork1289]

@support_huihui nvfp4 would be nice 😊

[Viktor @viktork1289]

Tic Tak Toe - Qwen3.6 35B 4bit vs. Laguna S2.1 nvfp4

Qwen wins 12 - 7

[Viktor @viktork1289]

@BosonJoe I power Off mine when I don’t use it

[Viktor @viktork1289]

@sudoingX I only get about 13 tok/s on my spark 🤔

[Viktor @viktork1289]

@MosiMoradian Have fun with it, got my Acer gx10 last week and I love it.

[Viktor @viktork1289]

@ivanfioravanti Thanks :)

[Ivan Fioravanti ᯅ @ivanfioravanti]

I love the new web UI that @viktork1289 added to the LLM Context Benchmark project!
It simply runs, compares and exports data out of it!
Endpoints tracking feature is top! Great job!

There is a docker version, but you loose the native MLX engine testing, you can test only remote endpoints.

[Viktor @viktork1289]

DGX Spark vs Apple M4 Pro
Qwen3.6-35B-A3B NVFP4 vs MLX

Wasn’t able to run the 256k context test on 64GB Memory on the Mac Mini.

[Viktor @viktork1289]

@CristianoRevil @MiaAI_lab @ivanfioravanti @mr_r0b0t @SpaceTimeViking @Tech2Wild @AtlasInference Depends on the use case. When it’s a big model and I use it only for chatting I’m fine with 10tok/s. For coding and agentic usage 40tok/s+

[Viktor @viktork1289]

@ivanfioravanti made a few improvements to the webui for the context benchmark tool :)

[Viktor @viktork1289]

Hello Little Baby @NVIDIAAI