Jigs

Just got 44K+ monthly views on YouTube with 700+ subs About to hit 1K soon. DMs are open if you want to sponsor a video 😉

Just tested Tencent HY4 and used it to create a 3D voxel art scene. The model is definitely better than the preview version. Is HY4 really GPT 6.1 Sol level?

just tested Crouton Alpha free on DeepSeek Harness. It generated an impressive 3D aquarium at 371 tok/s, and the result actually looks really good. Could this be a frontier level model?

Just compared LongCat 2.5 with Claude Sonnet 5.5 (Medium). And Sonnet clearly mogged LongCat 💀 Sonnet generated the code much faster, and the output quality was noticeably better too. Are open models started falling behind?

Just completed the first milestone on YouTube after working on this for almost a year. Let's see if I can hit the second milestone before the end of this year.

Just tested Qwen 3.5 4B 3D Coder on 2x T4 GPUs. It was running at 40-200 tok/s, but couldn't even build a proper demo. Even with clear instructions from Claude and multiple iterations, it kept failing.

Just tried running Qwen 3.5 4B 3D Coder locally, but it barely managed 3 tok/s. At that speed, I can't even build anything with it. Running models locally isn't for everyone 💀

Just tried Pika H3 Flex on a 16GB GPU, and it generated a 480p 5 second video in just 12 minutes 💀 For comparison, MiniMax H3 used to take around 30 minutes with the 4 step LoRA, and Pika H3 Flex seems to deliver even better quality. That's a pretty impressive improvement!

Just ran DeepSeek V4.1 Flash on an L4 GPU using Strata. Got 2.34 tok/s. Yeah, it's pretty slow. But running a 763B parameter model on a GPU with just 24GB VRAM is still wild.

Just published 2 repos for running AI models on Kaggle! Wan2GP: Run image, video, TTS, and other types of models. github.com/jigs10/wan2gp-kag… Unsloth Studio: Run text generation models. github.com/jigs10/unsloth-st… Both notebooks store models in the /tmp directory. Hope this helps anyone experimenting with AI models on Kaggle!

Just tested MOSS Image on 2x T4 GPUs on Kaggle. It generated this image at 480p for free in just 3 minutes and 10 seconds. Is it better than Qwen Image 2.1?

Just tested Kimi K3.1 Pro and it built this website. Cost me around $0.32, and honestly, it doesn't look like typical AI slop. Do you think this model is actually good or not?

I tested Kimi K3.1 Pro against Claude Sonnet 5.5 (Medium). Sonnet 5.5 performed much better in my tests, despite their similar API pricing. Is Kimi actually worth it?