About 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.
Over 3 million 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
Recent supporters

