Ivan Fioravanti ᯅ

💎Tale of o3-mini-high and a Lonely Diamond Shader.

This has been not easy, a lot of back and forth between me and the model, but the fast response times + reasoning + coding skills made this feasible!

R1, O1... maybe they can, but this would have taken forever!

1/🧵 

Generated code is available on ShaderToy directly if you want to play with it.

Initial prompt was: "Create a realistic effect of a diamond shader using glsl code that runs in shadertoy, use cubemap loaded in ichannel0 as background to be reflected in the diamond. Create a complex and beautiful diamond shape. Pay attention to details and ensure rotation of camera shows the beauty of the shape and effect."

2/🧵

https://www.shadertoy.com/view/43VfRc

Initially the model was insisting with low number of facets 4 and was not adding more than that 🤷‍♂️
It has bee a fight/collaboration between Human Brain and Artificial Intelligence 😂

3/🧵

Here one of the first working version I got from o3-mini-high.

4/🧵 

DeepSeek-R1-0528-5bit on MLX pushing M3 Ultra 512GB to its limits! 501GB used mem visibile on mactop in the video!

Context: 4K tokens
Prompt: 190.29 t/s
Gen: 11.37 t/s
Peak Mem: 487.48 GB!

THIS IS APPLE MLX! 

Here you can find the quantized (5.5 bits) model on Huggingface: https://huggingface.co/mlx-community/DeepSeek-R1-0528-5bit

In the video I left last part normal speed to give real idea.
Thanks Apple MLX Team for being able to do something like this! 🙏

8K context achieved! 💪
Prompt: 8145 tokens, 167.260 tokens-per-sec
Generation: 1101 tokens, 10.158 tokens-per-sec
Peak memory: 496.840 GB

16K!!! MLX no limit! 💪💪
Prompt: 15777 tokens, 131.764 tokens-per-sec
Generation: 1265 tokens, 7.329 tokens-per-sec
Peak memory: 510.726 GB

Here the video of a second try with 16K. 
DeepSeek-R1-0528-5bit running on MacOS 26 Beta 2 with mlx-lm 0.25.3 🚀 