DeepSeek v4 Flash in one DGX Spark in action

How big is the context window in that demo? :)

Take a guess :)

262k? Or the full 1mil? Either way awesome

How does the intelligence compare to 2xGB10 setup?

A bit less but still excellent

So, was my purchase of a 2nd GB10 (just about to be dispatched) a waste of money then?

No, 2 units is still much better. MUCH more context and concurrencies and still faster

Ah good, thanks, because I've just had a notification it has been shipped :)

Hey @MiaAI_lab do you clock your DGX or go full steam?

Clock it

Mhz?

sudo nvidia-smi -lgc 0,2200

Is this the deepseek harness? How is it?

Yes. My favorite harness for now outside of Hermes agent

need to test this!

do it it's really awesome

seems it is only web UI?

only web yes

fitting v4 flash on a single spark is wild tbh

When you do testing like this do you have multiple computer setups to leave your glm 5.3 running or do you have to keep model swapping to do this?

this is gonna be my weekend project excited

Very excited for the recipe

Thanks, coming soon

in one ??? 🤯

yes 😊

🔥🔥🔥

👏👏👏

Can you squeeze it into one 96 gb in a useful way? that would be the dream

t/s?

~47 tok/s structural

Wb 100 tok/s on a MacBook 24 gb of ram w 95% quality. Hypothetically

Can’t wait to see a demo on a single 5090

Demo ds4f on a 5090? 😬

how accurate or the quality for the quant?

Excellent quality!

WHAT wow that’s faster than i thought it would be… time to stock up on sparks i guess

!!

Waiting for it! Super excited!

That’s not fair. You’re teasing us now. Lol

Trying to deliver the best product I can before releasing it

You’re the best, appreciated.

It’s moving

not bad at all! very usable

Very!

What’s your reasoning level pick here @MiaAI_lab ?

What harness is this? It looks good.

@JensenHuang can i have one? 🥹

How much context do you get on this config?

@MiaAI_lab how can this be faster or even as fast as the two DGX version ? So it means the new version of the cluster will be faster ?

One DGX Spark running this, solid setup

what do you think about deepseek harness?

So fast💨

Are you using dwarfstar? or something else?

Still prefer glm 5.2 across 3 instead of flash on 2

What quant? This has me bout to go to microcenter tonight

that's the real unlock for large language models not just scaling compute, but making it deployable on existing hardware, DGX spark included

MiaAI-Lab made my dsv4f 0731 journey very smoooooth!

Whoever u r. Thank you for your guidance on running stuff locally. Do you have any articles about walk thrus and stuff. Wish u did