Tencent Hy-MT2 is now under Apache License 2.0

Posted by sword-in-stone@reddit | LocalLLaMA | View on Reddit | 15 comments

nice update bois

[-]

crantob@reddit

Just tried the 1.8B-Q8_0. 13.1t/s on my stock llama-cli on ryzen 3000u laptop.

Taking just 440mb, this could run on an Omap 4 phone from 2011. Sure it'd be 0.5t/s, but it'd be enough for high-value texts. Babelfish is here. Great day.

[-]

fragilesleep@reddit

I just ran some tests on HY-MT2-7B-Q6_K.gguf with all my usual Japanese visual novel examples... And wow, this is fantastic, by far the best local model I've tried for this task.

Now I probably should delete hundreds of GB of other older models I had.

[-]

Embarrassed_Soup_279@reddit

did you find this model to be better than gemma 4 series?

[-]

fragilesleep@reddit

I think it's a bit better than E4B with thinking enabled (and much faster, of course), but I didn't test with the bigger Gemma 4 models, since I need this to be small to be running while gaming.

[-]