Unterstütze Laurent Zuijdwijk

I build open-source quants and kernels so big models run on hardware people actually own. Gyro puts Qwen3.8-Flash-Next (126B) on a single 32 GB card, and on 12 GB VRAM plus RAM, and it runs on llama.cpp and Strata. My 27B quants are measured against Unsloth and ISTA at matched file sizes, with the numbers on each model card. Over 800k downloads so far.

Most of this is built on a couple of Strix Halo boxes at home. Large models need rented GPUs, and each model takes about a week of work.

Current goal: $600 for the next large-model encode (about 3 days on a rented multi-GPU box). I’ll post what was spent and what came out of it.

Models: huggingface.co/agentionai · X: @laurent_zw