AmenoTurbo (∎, ∆)

Local LLM Enjoyer

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@astekz Norway beat england

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@slavamorozov23 @ZixuanLi_ @TencentHunyuan None of the labs release models that run on mainstream hardware, only qwen 3.5 and gemma.
Deepseek, GLM, Minimax, Kimi, Mimo, etc..

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@slavamorozov23 @ZixuanLi_ @TencentHunyuan I have enough vram to run hy3 but don’t have enough for deepseek v4 pro.

So yeah their models looks very good and looks way better than deepseekv4 flash for the moment.

(Not saying deepseek won’t release a better model soon.)

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@hebbarmp @ZixuanLi_ @TencentHunyuan 28

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@slavamorozov23 @ZixuanLi_ @TencentHunyuan Deepseek is a 1.6T model, hy3 is a 295B

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@jietang @Zai_org I think either a faster model or more optimized thinking so that it is shorter

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@zanoga Thank you for the insight it, many of us are probably half way through this path

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@zanoga I am curious what kind of risers do you use ?

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@pipenetwork Can you make a GGUF of the 25% Reap ? 😁

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@LLCoolChris_ En dessous de 40 c’est limite selon moi

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@pipenetwork @huggingface Do you consider a REAP 25 IQ2 XXS?

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@0xSero Honestly I think REAP in its current state is way worse than quantizing, however as you said if no one tries, there no way to make it work.
I actually root for your success but I am convinced that right now REAP does not work, would love to be proven wrong. 🤣

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@NexEcosystem Btw thanks for the model m, I am using this one at home and it passed my go to test successfully when qwen 3.5 397B and Minimax M3 didn’t

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@Prefeitura_Rio congrats !

https://t.co/Ffse8kPbjR

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@sudoingX I am confident they will ship open weights just wait

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@iForgor_PM @atteonthephone Actually without reading the rules I would say 0 doesn’t count jut after reading it should count 😂

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@HadrienClouet La question c’est comment les gens peuvent trouver le bilan de Macron satisfaisant

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@Seb_Gode @ctnzr Yes I use this one but there is no version for vllm

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@pierrelezan @ctnzr Yes, but the persons is thinking about buying multiple 3090 used, he will not buy a 5090 because it cost too much and will likely buy a MAC
Many ppl go for a MAC for this reason.
If a user buy and use 3090, they learn the NVIDIA experience and maybe will use it at work on DC GPU

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@ctnzr NVIDIA ultimately makes money and benefits from home AI but not supporting « old » hardware is kinda bad advertising for future buyer.
Also it wouldn’t cost much to support older hardware and would mean a lot.
2/2

[AmenoTurbo (∎, ∆) @AmenoTurboAtmo]

@ctnzr I am surprised nvidia doesn’t release models for other older hardware, I am on multiple 3090 and often find it hard to run a the latest models. For exemple MiMo 2.5 or Hy3 preview have no awq 4bit options yet but NVFP4 exists and sometimes from NVIDIA itself.

1/2