And to remind you of bipolar nature of my mind, here are my notes on local LLMs:
1. 99% of them are horrible, just outright barely usable even for low-risk applications. Qwen 3.6 derivatives passed some benchmarks but SUX ARSE (Prism 27B included). Nvidia's own stuff is trash too.
2. The exact model is more important than the family. The size (8B, 12-14B, 27B, etc) maters less than how tuned and mixed it is.
3. For my goals I ended up with two very capable models:
https://huggingface.co/Brunobkr/OFFELLIA_Gemma-4-E4B-8b-Claude-4.6-Opus-Reasoning-MTP - a tiny 7.5B model but very well tuned for tech stuff. Makes some mistakes, so ask wisely. Can run on a laptop.
https://huggingface.co/mudler/gemma-4-26B-A4B-it-Claude-Opus-Distill-APEX-GGUF - fairly big MoE for all sorts of stuff. Eats tons of RAM, but Mix of Experts with APEX allows it to still be very fast.
And I run stuff via LM Studio
#AI@GameDEV