Major FAFO

@elonmusk Peaceful Easter!

✝️ Happy Easter! ✝️

27b still King an Gemma 4 31b a Queen, test them both, start with ggml or unsloth quants, some other releases may not be so accurate. I think the best way to use 2x3090's is to use the NvLink connector, if you find one. pci-e will kill your bandwidth between them, risers: massacre.

Im doing the typical gamer setup to local inference pipeline rn Learn from my upcoming posts What are pcie lanes, what’s x16/x8x/4, how slow is 2x3090 on prosumer mobos, all coming up. Just to see if qwen 27b is still the king.

@FLOTUS Peaceful Easter!

Happy Easter!

Peaceful Easter!

I wish everyone, including the haters and losers, a very happy Easter!

Stop normalizing saying "Happy Easter" sounds pretty harmless to me. Better to say "Peaceful Easter" - this is a powerful message. Thank you for your attention to this matter!

The tomb is empty. The promise is fulfilled. Through His sacrifice, we are redeemed. We stand firm in faith, courage, and truth—Happy Easter.

He is Risen.

@bridgemindai x.com/i/grok/share/e… consider this

Claude Code rate limited me so hard I bought a $5,000 NVIDIA DGX Spark. Arriving tomorrow. A personal AI supercomputer. Anthropic cut off OpenClaw users. Slashed Claude Opus 4.6 rate limits. Told $200/month Max plan customers to use less. Then gave us a credit as an apology. This is what happens when AI companies have too much power over your workflow. One update and your entire stack breaks. Local models are the only infrastructure no one can throttle. No rate limits. No 529 errors. No surprise policy changes. Tomorrow I'm testing the DGX Spark live on stream. Running local models through real vibe coding workflows. The goal is simple. Never depend on a single provider again.

@eracoon @YourAnonOne beat me to it, I would have farted my way out of this

@YourAnonOne That’s what farts are made for

An isolated astronaut in weightless can remain stuck forever if he cannot reach a surface to propel himself

Bros pcie4 x4 is bad

@nikitabier @nypost they probably were right, bloody bastards may be on to something after all

@nypost Can I say something without everyone getting mad

Notorious Gen. Soleimani's sultry grandniece led lavish lifestyle touring US hotspots, as her mom promoted Iranian regime trib.al/y38evjw

@loktar00 llama.cpp is great with its webUI, only needs a HF model downloader and a simple UI to set loading parameters

HuggingFace building a local model provisioner on top of llama.cpp.... this is going to make Ollama completely irrelevant. Harbor already does this but having HF behind it means model compatibility day one.

@LottoLabs Check with nvlink and report back please, should keep the speed similar while doubling the VRAM

I just bought another 3090 don’t listen to this guy

@TheAhmadOsman Did you tested them in BF16?

- Gemma 4 26B-A4B > Qwen 3.5 35B-A3B - Qwen 3.5 27B > Gemma 4 31B

@MaziyarPanahi Cool, what was the setup methodology?

Gemma 4 watches raw video. Understands the scene. Then prompts SAM 3 to segment and RF-DETR to track. One AI directing two others. Fighter jets. Crowds. Aerial defense footage. All three models running locally on a MacBook. No cloud. What scene should I point this at next?

@stevibe Ere you performing the tests in BF16?

Qwen3.5 27B vs Gemma4 31B | Canvas Creativity Test Why HTML Canvas? Two reasons: 1. It's unforgiving, one small mistake and the whole thing breaks 2. We kept prompts short to test real creativity, not instruction following 4 rounds: - Analog Clock - Hyperspace Tunnel - Growing Tree - Black Hole Both nailed the clock, but the other three is where it gets interesting. Looking forward to Qwen3.6 open-weight release!

@sudoingX always been @grok thoughts?

wooo grok is different now

@julien_c Qwen3.5-27b as a good all-a rounder, reasons well don't go below Q4, uses about 350W on 3090, Gemma4-31B-it is fine but trained for benchmaxing, it memorised the answers, then fails if you change variables

so…. Qwen3.5 or Gemma 4?

@hallerite @ggerganov 15 tok/s on Q4, both tests with full context

@hallerite you're right, running gemma 4 on llama.cpp CPU better speed on CPU than GPU (6 vs 1.5 tok/s) in VRAM restricted scenario. I remember long time ago @ggerganov said all you need is CPU for inference, I think he saw it early

you can tell that gemma4 was not trained on nvidia gpus