Valeri Yehezkely

Valery Yechzkeli | AI dev
Local LLMs · LangGraph multi-agent systems
6 yrs in hi-tech · Going indie

[Valeri Yehezkely @ValeryXTech]

@germanburgardt i will pay 200$ to play a game like that even if it only be a clone of DBZ

[Valeri Yehezkely @ValeryXTech]

@GrilliotTodd is this made with AI? This is crazy!

I used to use PixelLab, but it did a bad job with monsters, like a crab with 6 legs

[Valeri Yehezkely @ValeryXTech]

@MiaAI_lab can it run on a single 5090? And if yes, is it better than 3.8 27B Q5?

[Valeri Yehezkely @ValeryXTech]

@theindiescrolls This looks so good with es6 out or without you need to release this

[Valeri Yehezkely @ValeryXTech]

@benchmark_lb900 Actually, this looks very good compared to the price, I guess. How much did the Kimi Mario cost you?

[Valeri Yehezkely @ValeryXTech]

@urieli17 כנירא שבעתיד זו תהיה ההפרדה בין המעמד העשיר לעני מעמד המודלים הזולים מעמד המודלי ם היקרים

[Valeri Yehezkely @ValeryXTech]

@AiBattle_ Only 1% how much do you pay for the sub?

[Valeri Yehezkely @ValeryXTech]

@LexmountAI looks useful thanks

[Valeri Yehezkely @ValeryXTech]

@AIMind_Ai On https://t.co/ubLTs6CxbT

[Valeri Yehezkely @ValeryXTech]

@AIMind_Ai Who do you rent it for 500$? I have pro 6000 rtx 96 gb and 128gb ram ddr5 and I barely earn 12$ a day

[Valeri Yehezkely @ValeryXTech]

One kept fixing its bugs until everything worked. The other fell into a pit and died. That was the whole difference.

[Valeri Yehezkely @ValeryXTech]

Ran two local models on my 5090. Qwen 3.6 vs Qwen 3.8. Same 4 game briefs, no human help.

https://t.co/T6Wq0qzPVM

[Valeri Yehezkely @ValeryXTech]

Final: Qwen 3.8 8.5/10, 4/4 games playable.
Qwen 3.6 4.9/10, 1/4 playable.

[Valeri Yehezkely @ValeryXTech]

e. For real projects, the self-verification loop is the deciding factor. That is what "better model" means in practice.

Full breakdown + playable builds:
https://t.co/T6Wq0qzPVM

[Valeri Yehezkely @ValeryXTech]

3.6's Mario 2D: 5 known bugs, two game-breaking (jump too low, finish line unreachable). Its 3D Mario jumped 2.88 units; the platform sat at 3.5.

3.8 kept testing and fixing until the brief was met. Coyote time + jump buffer + variable jump when asked for one.

[Valeri Yehezkely @ValeryXTech]

@sgl_project @Alibaba_Qwen You mean 203 tps with any context
Because I tried it with my 5090 with 160k context and it's 50 to 90 tps

[Valeri Yehezkely @ValeryXTech]

@MiaAI_lab @UnslothAI What you recommend on 5090
Q6 or nvfp4?

[Valeri Yehezkely @ValeryXTech]

https://t.co/aTwSnwk0q5

[Valeri Yehezkely @ValeryXTech]

First MiniMax H3 video, ran it locally on my 5090.

It came out really good, especially the combat scenes. I'm not a video person, this is one of the few times I've made something like this.

Prompt + config in the comments. Tell me if I messed something up.

[Valeri Yehezkely @ValeryXTech]

Models: MiniMax H3 UNet (int8) + Qwen3-VL 32B CLIP (NVFP4) + H3 Video/Audio VAEs

Happy to answer questions about the workflow.

[Valeri Yehezkely @ValeryXTech]

iMax H3 · 16:9 · 0.9MP · 15s · seed 563400907186515 · RTX 5090 32GB · ComfyUI