SSH

i make, like, use, tune AI

[SSH @SSHCodes]

@TheZachMueller 60b-120b, the best model sizes

[SSH @SSHCodes]

@HobermanSpencer @sama fast chip good but need millions of dollars to get

[SSH @SSHCodes]

@Tech2Wild @Tnimbus it’ll be extremely tight on one spark. It’s 125b a6b but with a 51b engram and because the spark has no dedicated CPU ram it has to live in the same spot so even at NVFP4 it’ll almost hit the limit

[SSH @SSHCodes]

@Sergizzzz4 @superalesha probably not

[SSH @SSHCodes]

@QuixiAI the new m5 ultra seems a lot better

[SSH @SSHCodes]

"farris" on lm arena seems really good at design

[SSH @SSHCodes]

im excited for qwen 3.8 flash next, one of the major issues I have with qwen 3.8 27b that it’s bad at knowledge stuff and this will probably fix that

[SSH @SSHCodes]

@Tech2Wild gonna be very tight on 1 spark

[SSH @SSHCodes]

@jun_song an actual PC is better suited for this model, VRAM holds the model and RAM holds the 51b engram

[SSH @SSHCodes]

@v_oleksiienko @arena @Alibaba_Qwen There normally isn’t any cost per task lol this model is meant to be ran locally

[SSH @SSHCodes]

@piotryordanov @SergiiioBS 125b a6b but with a 51b Engram that uses less compute than the model but sits in memory to give it more capacity

[SSH @SSHCodes]

cool iOS app I made that lets me fully use my pro 6000 securely from anywhere that sol made for me and generate videos, music, chat with qwen, etc.
if you like it i will open source

[SSH @SSHCodes]

@Maciej25956571 @MiaAI_lab @sethforprivacy it’ll fit with nvfp4 with very little room especially with big kvcache

[SSH @SSHCodes]

@MiaAI_lab @sethforprivacy on only one spark though it will have very little breathing room with its huge 51b egram table

[SSH @SSHCodes]

Qwen3.8-Flash-Next might not run well on one dgx spark at all 😨
it has a huge 51B Engram table that even when compressed to 4 bit would use around 25gb of RAM, but because of DGX spark unified memory the total the entire model would use 105–125GB of unified memory total making it very close

[SSH @SSHCodes]

@danieltvela 300tok/s

[SSH @SSHCodes]

@ItsmeAjayKV i use sglang recently, seems faster (don't have rtx 3090 tho)

[SSH @SSHCodes]

@qinzytech where does glm score? if glm scores horribly on this too could be another tiny sign

[SSH @SSHCodes]

@blueemi99 I tried grok bot, found no use in it, and never used it again. other agents just already do everything I need

[SSH @SSHCodes]

@NavDoesTech @bridgemindai @poteto man idk if even the strongest skill csn fix opus mess