gguf when??? cmon guys, where is bartowski?!
💀
Anyone willing to donate around 20TB of SSDs to my cause?
It's "just" 4.6 trillion params, It should take you seconds, barely :D
Hi! We will quantize this ourselves, because of the architecture.
Anyone willing to donate around 20TB of SSDs to my cause?
It's 7TB, actually:
Yup.
Even using the experimental https://github.com/ggml-org/llama.cpp/pull/14810 to go from fp8 safetensors -> bf16 GGUF you'll need 7.72+7.72*2 = ~23+ GiB just to get started, let alone space to store any quants (or delete the safetensors and hope you don't need them again hah)
(or delete the safetensors and hope you don't need them again hah)
I thought that'd it was possible to quantize one file at a time
Look guys! New Reflection-Llama-3.1-70B just dropped. Blows will be traded with GPT5.
Right actually your probably need more like ~30tb at peak 14tb for the bf16 then 7tb for the Q8_0 and another 7tb to split the Q8_0
Just a protip: when @mattshumer tried this last year and people realized it is a scam, he created a site where you could test his model and it was routing requests to claude. This should buy you some time and distract people so you can escape and rebrand.
Hot single LLMs are in your area right now. All you need to do is click this link and pull out that credit card.
That's not the point. This was intentional, not a fake. Alpha was supposed to be this way. Mostly copied. The gold is in the dynamoe software. The rest is just rock. You will see once dynamoe comes how it all comes together.
Till then please be patient.
Till then please be patient.
Hey at least you know how to do this part of a scam, but you really shouldn't be doing that to the negative people like me. You use this on people who think it is a mistake and something is wrong. The best course of action for people like us is to either ignore or violently ban and censor everything.
I will be cheering you on man. One day you are gonna be a great scammer. You just need to keep at it and get more practice.
BTW someone from our secret discord ran benchmarks for you. Feel free to add this to your card.
Why would you change the license when you're using InternVL weights, which is apache2?
You use this on people who think it is a mistake
I told you this isn't a mistake. This was the plan. If this feels like spam, then that's our fault of not communicating this enough. We aren't trying to hide anything here. We're trying to show it.
Why would you change the license when you're using InternVL weights, which is apache2?
Because we're also releasing 2.5 is the background so we don't want it bogged down by 3.
Because we're also releasing 2.5 is the background so we don't want it bogged down by 3.
But this isn't about your models. It's just a crappy thing to embed someone else's literal weights and change the license.
It is crappy. It was meant to be. Until we have the real Deca 3
It is crappy. It was meant to be. Until we have the real Deca 3