Sabitlenmiş Tweet
Drop your GPU below and I’ll tell you exactly what model and config to run on it.
JOKES. No need.
Qwen 3.6 27b @UnslothAI . All day. All night.
English
BlackwellBoy
8.7K posts
BlackwellBoy
@Blackwellboy
Sovereign local AI architect • 1208 GB heterogeneous beast (RTX 5090 + 5×3090 + 8× DGX Spark + M4 Mac Mini) Hermes ftw.
Katılım Ekim 2012
1.8K Takip Edilen2.1K Takipçiler
@AgenticOra been paying 800-900 in aus which I think is 600+ USD?
English
HOT TIP 🔥 - Facebook marketplace, search gaming PCs, not just GPUs. Blokes selling their whole rig don’t know the 3090 inside is worth more than what they’re asking for the lot. Been sweeping my area clean for weeks. Comes with bunch of possible useful extras too.
English
@Teknium @jangofe83800676 Damn, sorry to hear homie ! Head up! Keep on keeping on! Much love. ❤️
English
Hey everyone. I haven't been very responsive on here the last week.
My dog, Link, who I've raised since he was a puppy over the last 13 years, passed away yesterday after being in the vet ER's ICU since last Wednesday for heart failure.
I put together some of my favorite pics of him to share so you all can see the most awesome animal friend I could ask for.
I'll be a bit slow probably through this week too, hope you all can understand 🙏
English
@Blackwellboy so you're why i can't get a replacement gaming pc huh?
English
@BosonJoe Depends what country you’re in probably different to me
English
🛸GIANT ALGO LURKS NEAR JUPITER🛸
These images captured on July 8, 2026 by a telescope show something impossible to explain orbiting Jupiter.
It's not an asteroid. It's not a camera glitch.
It's a massive structure with a shape that defies every known natural law.
A mothership? A UFO of cosmic proportions?
Space agencies are maintaining absolute silence, as always.
What is that technology doing at the edge of our solar system? Are they watching us... or preparing something worse?
The proof is right there. Before your very eyes🤫
#ufos #disclosure
English
The new Miss Italy is an African dummy.
The beauty pageant brought to you by Netflix
English
there is a moment in a mans life where he stops counting in gigabytes. 1.2TB achieved. no support group. no intervention. just me and the fabric.
BlackwellBoy@Blackwellboy
soon 1TB of VRAM. the scary part isn't the number, it's that nobody stops you. my feed sees me at 824GB and whispers "1 more Spark." so u accumulate. my power company sends handwritten Xmas cards now. still waiting on my first Sparks Anonymous meeting. attendance: me. @NVIDIAAI
English
@Blackwellboy Yep. Everyday there are new local model recipes I have to try, and I'm still waiting for some new models!
English
Local AI chasing, waiting for that qwen 3.7 or 4 whatever they decide to give us next
George Pu@TheGeorgePu
Are most of us these days just frontier chasing? Whichever lab has the best model, we go for it. The moment someone else catches up, we switch. Billions of dollars for just a few weeks of moat. What an era to live in.
English
@Blackwellboy I don't understand how people are still using Claude.
It has become Russia 1917.
English
Asked Fable 5 to help with a bash script. Flagged.
"The safeguards are intentionally broad right now"
Mate I'm renaming a folder, not enriching uranium.
Most annoying thing in AI right now.
English
English
BOOOM!
Like is said frontier model-like running one a typical laptop!
New approaches are not in the mine of the “that’s impossible” folks.
744B GLM-5.2 on a laptop, not just possible but running quite nicely.
Garage Open Source always wins.
You good Dario?
Zane Chen@chenzeling4
744B parameters. On a laptop. With 25GB RAM. Colibri runs GLM-5.2 (744B MoE) in pure C with zero dependencies. The trick: only ~40B params activate per token, so it keeps the dense part resident and streams experts from disk on demand. A single 2,400-line C file. No GPU, no BLAS, no Python at runtime. This shouldn't work. But it does. ⭐ 2.1K #AI #OpenSource github.com/JustVugg/colib… Follow for daily dev finds 🔔
English
@Blackwellboy 220 is my choice
x.com/superalesha/st…
Alexey Fateev@superalesha
Capped each of my 4x3090 from 220W to 300W and re-ran Gemma 4 31B (AWQ, vLLM, TP4, single stream). Decode: 73 -> 77 -> 78 tok/s. That is +37% board power for +7% more speed. The curve goes flat after 220W. Stock power is already most of what these cards give you.
English
Your 3090 does ~90% of its inference throughput at ~50% of the power.
⚡ 320W → 114.8 tok/s
⚡ 225W → 105.1 tok/s
That last 95W buys you 9 tok/s and 12 extra degrees.
One command: nvidia-smi -pl 225
Run your own sweep. Every card has a sweet spot. 🎯
English
@u1tra_instinct doing amazing work out here again, taming some of the most wild beasts in the game. His DS4V was brilliant, can’t wait to suss this ! ✅
keys 🧪@u1tra_instinct
For those who are not able to obtain/run the Full GLM 5.2 quantrio INT4/INT8 Mixed - Abliterated. Here is a clear instruction and information on how this is done and get you setup and running. GIthub Repo updated for one-shot script (see link in Article).
English