SSH

i make, like, use, tune AI

[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

[SSH @SSHCodes]

@maria_rckk where’s the one and only Maria Rock tho

[SSH @SSHCodes]

@curiousbigfish @maria_rcks that doesn’t take away the huge token inefficiencies and price tho

[SSH @SSHCodes]

@NavDoesTech am not over clocking to risk stability sorry

[SSH @SSHCodes]

RTX pro 6000 is here, already pushing 250 tokens a second on qwen 3.8 27b 🤗 many models to come

[SSH @SSHCodes]

@maria_rcks fair lol it is extremely token inefficient making it extremely expensive with intelligence literally worse than Gemini somehow on AA which is why I think it belongs down there with Google

[SSH @SSHCodes]

@DR4G4NS @blueemi99 apparently its horrible at front end but that isn’t what I was using it for

[SSH @SSHCodes]

@MiaAI_lab @bot minimax H3 is very slow I like LTX 2.5 better for its speed

[SSH @SSHCodes]

@Anbeeld too high for the tier list, qwen too good for it

[SSH @SSHCodes]

im not everyone but I THINK i nailed this let me know what you think

[bluedev @blueemi99]

If anyone wants to make their own tier list:

https://t.co/VfO28PdrP5

[SSH @SSHCodes]

@blueemi99 fair lol

[SSH @SSHCodes]

“quietly deleted their posts” meanwhile them apologizing and publicly saying they deleted it lmao. This guy sketches me out, constantly hating on people and glazing Elon like what?

[NIK @ns123abc]

OpenAI patched the glitch and theo from t3gg and his intern quietly deleted their posts so they don’t get blamed lmao

[SSH @SSHCodes]

@blueemi99 1. yes
2. fair enough
3. can be dumb sometimes but for me it was mostly good which is why for it I think c tier is fair

[SSH @SSHCodes]

why do people think the dgx spark is better than the pro 6000? IMO even 4 DGX sparks aren’t as good, I do training and the memory bandwidth and worse compute kills the DGX spark for me