llmwish.open

GLM-5.2 on 8× DGX Spark — v18 update 🚀 Jump from 1,200 → 1,329 t/s prefill Peak decode hits 66 t/s (was 54) Repo with both v16 and v18 builds: github.com/ciprianveg/gb1… Forum post with benchmark: forums.developer.nvidia.com/t/glm-5-2-int4…

@HusamSoboh Sigur

@ciprianveg Salut, am putea sa intram in contact? As avea nevoie de niște sfaturi.

Preparing 16xGB10 cluster for Kimi K3, Deepseek V4 Pro, and future GLM 5.5 and Minimax M4..

@LD_Theron mikrotik.com/product/crs804… can connect 4, 8 or 16 depending on the breakout cable being used. 400to2x200 or 400to4x100

@ciprianveg Dude, it's you I was looking for. How did you set up your system? I want to do the same, but I keep hearing about switches and bandwidth, etc., and I'm stuck. I can only connect 2 out of 8.

Kimi K3 Coder Reap on 8xGB10 - 32t/s avg gen speed in coding task. This is while waiting for the full 16xgb10 setup to run the full k3. Considering that this one has same no of active experts, it is a good speed for my first, unoptimized, run of KIMI-K3-CODER-REAP-320-MXFP4 with dspark. I will work on improving these results and let you know.

@no_stp_on_snek It exists

@ciprianveg Dedicated AC for that room needed

@DawsonBotsford 1-1.7kwh

@ciprianveg What’s the energy usage like? Seems like mounting some 6000’s in a case would be cheaper and run on a normal circuit

@u1tra_instinct Awesome!

🚨🚨🚨🚨🚨: most requested since the DeepSeekV4-Flash GA 0731 release yesterday. Now Abliterated 32/32 100% compatible with DSpark please if you like my work and want to contribute please can help via X-money, GoFundMe link in profile 🙏 any little bit helps, whether for token credits, coffee, time, power, hardware. HF repo: huggingface.co/drowzeys/keys-…

@samsaied55 Ray

@ciprianveg Are you using EXO by any chance ? How are you handling distribution

@hugoscode Cca 70k

@ciprianveg How much was this setup?

@tedixh1 Price and better cooling

@ciprianveg Why did you pick the GB10's over the founder editions?

@0xSero So true..

You are not alone We are not separate My daughter's shoulders are my shoulders My son's hands my hands My wife's heart my own heart

@antoniodearmas What is fast enough for you? 50tps+ single stream in coding for glm5.2 is good enough? For me, yes. Multiple requests are also scaling well, till 6-8.

@antoniodearmas Owning AI instead of being owned by it

DeepSeek V4 Flash 0731 @ 118.2 tok/s First trial. Just getting started. Going to see what I can push this to without sacrificing quality

@tarat_211 Cca 70k 16 asus, switch and cables

@tarat_211 6 Cca 70k

@dev_null321 Old love..

@ciprianveg lol why the raspberry pi mouse and keyboard ?

@defiobi Look at this thread forums.developer.nvidia.com/t/dual-spark-d…

@ciprianveg Where did you buy the case? Got 4 x gb10 i want to stack like that

@JamesMcPherson CRS804 DDQ with 4x 400to4x100 cables

@ciprianveg This is amazing. What switch are you using?

@u1tra_instinct AI must be local, even frontier level ones🥂

i'm in awe.... 16 DGX-Spark/GB10's cluster is being created here.. First one i have seen. you're not following @ciprianveg and watch what he is building frontier level local consumer inference monster, you might be missing out on the future when 1, 2, 3T parameter is locked out for you