Qwen3.8 Flash on a single DGX Spark. Smooth experience!

Loaded it today on my spark. Loving it so far

which app is this

My own, unreleased yet

wooow that speed is good! I've been playing with a variation of but this seems a lot faster. any repo to share? 😇

Pinned post

👀👀👀

I feel like my speeds are still a notch slower than what you are getting. What could be causing this?

32 on prose?

I’ll do more tests tomorrow, will report back 🫡

NP. I always have plans to improve. I mean REALLY improve.

👀

Solid performance

How does it compare to your deepseek-v4?

For a single spark this is a better experience

Which app is that?

My harness, unreleased yet

🥲🥲🥲🥲

How many tk / s ? Are you trying this model /dgx with more than one concurrency?

How would you run this on apple silicon Like m2 studio at 192gb

This is based on NVFP4, will not work on Mac Can't wait to have a Mac to create Mac repos

Started to build bare metal and software optimized builds for my m2 ultra 192gb using fable 5.1 and Astra Q went from 4bit to 8bit so far 8+ Threads at max native context Delivering 50+ decode Tok/s per Thread so far. Lots of improvements still to be

Mia: write a song Qwen: bet

has ablit plan?

Getting good performance on my Strix Halo as well, but not this good.

toks/s?

~35 tok/s prose!

Mia is there one for 128 MacBook 🙂🤞

I don't have a Mac (yet) ...

ok thank you 😌

Sweet, my dgx spark arrived today

local is the only version of agents where the bill doesn't scale with how chatty they get. that's the real selling point here, not the speed

That's cool! Could you explain where you installed it?

Single GPU local inference getting this smooth is seriously impressive🔥

Looks great!

Thanks because without nvfp4 it's so slow 😅

I don't know. Four or five lanes going in production, even lower context, acts a lot like GLM, which is to say breaks down, and needs a lot of maintenance of memory eating drift.

Put it on xhigh and show us

Ok

But what's it like at coding? I'm at the point of not minding buying a DGX Spark of even two. But the coding quality I've seen from Qwen3.8-27B and Qwen3-Coder-30B Q4KM on my RTX3090 is simply atrocious.

Does this run the ngram table off the ssd? Cause the nvfp4 model is 135gb, no? It can’t run in unified + you need room for the kv cache/context.

Creative writing at this speed with a single spark is unbelieavable... 27t/s.. What about structured/code, any difference? If not, then this new architecture must have something special.

Code is MUCH faster, like double the speed

YOO! so the reported 37tok/s is avg?

something like that yes :)

Thanks. This is great. it for long horizon task?

Would love to see it code. Doesn't give me much hope at this tokens/sec on just a random poem :)

epic, now I need something else to keep the other spark busy, lol