Live data from Hacker News

H3-metal – Native MiniMax-H3 inference for Apple Silicon

github.com

11–20 of 108 posts

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#13

wow antirez does not sleep

when you have enough money to not have to worry about anything, you can go back to your hobbies. in this case, his hobby is programming.

"enough money not to worry about anything" haha

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#14
post #4

How similar are Jeff Dean and Salvatore Sanfilippo?

My favorite Jeff Dean fact is that he’s also antirez. Which reminds me of my favorite Salvatore Sanfilippo fact. He’s also Jeff Dean

Is that what the identity function is?

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#17

wow antirez does not sleep

when you have enough money to not have to worry about anything, you can go back to your hobbies. in this case, his hobby is programming.

Being a world class talent is independent of financial situation.

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#18

Earlier quoted context omitted.

when you have enough money to not have to worry about anything, you can go back to your hobbies. in this case, his hobby is programming.

Being a world class talent is independent of financial situation.

talent w/o financial stability is a battery w/o circuit.

Re: H3-metal – Native MiniMax-H3 inference for Apple Silicon

#20

I've been using MiniMax H3 on my M5 Pro 64GB MacBook Pro through ComfyUI. It works extremely well. I had to modify the default ComfyUI workflows to use a GGUF quant (city96's ComfyUI-GGUF custom node, UnetLoaderGGUF in place of the stock loader) [0]. I use the model labeled Q5_K_M. There is Q8_0 available as well, which is 34GB and fits fine in 64GB unified memory if you keep resolution modest. The main issue is spee…

I wonder how much faster your m5 pro is compared to my M1 Max @ 64gb
Post reply on HN