@larpu_acc: lisa su (ceo of amd) was holding a tiny mini-pc (about the size of a router or lunchbox) during the presentation inside it there’s the ryzen ai max+ 395 (strix halo) with 128 gb of unified memory (shared for cpu + powerful integrated gpu + npu). on linux around 110 gb of that is available as vram (* ˘ ω ˘ *) this little guy (like the gmktec evo-x2) can fully run the 235-billion-parameter model (qwen3-235b, mixture-of-experts) and feels super comfy with deepseek v3 and other big models. no cloud, no separate graphics card needed! (* > ω < *) why this is exciting: • it’s the first x86 chip that can handle ~200+ billion parameters on a single die • according to amd, in some inference tasks it beats the rtx 5080 by several times (especially with huge models that are memory-bound) • such a mini-pc costs roughly $1400–2500 depending on config, and it pays for itself in just a few months instead of those $200–400+ monthly cloud bills (claude, chatgpt pro, cursor, etc.) in short: amd really leveled up mobile and mini platforms for local ai, and this is actually threatening part of nvidia’s cloud business and all those model subscriptions(* ^ ω ^) ♡ #amd #llm #opensource #software #larpochka

︎︎matts :

"you can now run one of the world's Ilm at home"

japo_tech_genius_pro_the_best :

holy larp

Lou.nfo :

boi would you rather pay a 20 dollar subscription to get 1-2T models or 1000 for a 200b shit model thats normally free on the cloud

fbi.gov :

larp doesn't what he's talking about

MurimPath :

What larp is this

Stevefox123 :

how about telling ppl it rundt so slow youll have to wait 30 min for a response if the task us large enough?

BadRally :

kimi k2.6 is a 1t model and is ~5 usd per million tokens

mikgazer :

It’s unfortunate to see the big technology companies slowly prioritizing AI and the business opportunities it could produce, but on a more positive note, we get to see cool innovations such as the 395 strix halo, beating a 5080 with that form factor is incredibly impressive.

wetware :

Or u can just ollama gpt-oss on ur spare GPU

Hackless :

Can it run GLM 5.2?

Fallor😴 :

AMD will always own nvidia when it comes to AI and LL

🇹🇷🇨🇾 :

Local ai is the future because it’s what people want

leafwayzz :

Larp overload

killua🎀 :

foxtheleague :

Why evb saying larp? Ion get it

Sndrdrx024 :

128gb is not a little guy

jebediah kerman :

larp

Dfrcyt :

I’m not broke so I’ll stick with Claude

Sing (Antarctic Singularity) :

tokens per second prob horrendous

️ :

correction, they made a mini pc that can run ai models for 4000 dollars

Purple :

I’d rather just pay for a claude/openai subscription

Mith :

So what's new? I did this last year in September—ran a DeepSeek-V2 235B model on an AMD Epyc 7282 (24 cores, 48 threads), 128 GB DDR4. TTFT was 3 min, 1.5–2.3 tk/s

darek :

fable 5 has a few trillion, 5 i think. and that tiny mini pc is worth like five thousand dollars.