anikifoss

/

GLM-4.5-HQ4_K

like 1

Hey @sousekd ! I downloaded GLM-4.6 and converted it to GGUF, but haven't gotter around to quantizing it yet. I've been having too much fun writing a minecraft clone with Deepseek-V3.1-Terminus.

Since there is demand, I'll run it overnight and post it sometime tomorrow.

Oh, thank you!
I read it as "hell yeah, 3.1 Terminus is goooood!" :)

I read it as "hell yeah, 3.1 Terminus is goooood!" :)

Yeah, totally! As of Deepseek-V3.1 release, all the coding tasks I wanted to do started magically working. And then Deepseek-V3.1-Terminus one-shotted code for a working WebGl-based minecraft clone after some manual pre-planning. None of the older models could get the 3D math and mesh chunking algorithms right, it was always a black screen or broken 3d models with incorrect texturing. But now Deepseek-V3.1-Terminus is spitting out fully functional games prototypes on demand.

In the Deepseek V3.1 release announcement (see " long context extension approach"), they mentioned more training for extended contexts. I think that made a huge difference, because now the model is able to work with larger codebase past 32k tokens, and even past 64k tokens. While previous models would degrade at 32k and become unusable after 64k tokens.

Sounds great. Do you use your HQ4_K quants, or something larger, to get those promising results?
Also, did you manage to make ik_llama handle tool calling correctly? I failed to get it to work last time I tried - so I fallback to llama.cpp when I need tool calling.
Sorry for all the questions, I haven't tried to use LLMs for the "serious" code generation yet, just small things here and there :).

I use HQ4_K quants exclusively.

I find both llama.cpp and ik_llama.cpp can be buggy in different ways, and it changes every month as old bugs are fixed and new features are added. I'm usually switching between them depending on the model, task, and runtime issues.