Running Frontier MoE Models on Consumer Hardware via vLLM-Moet
Deploy frontier MoE models like GLM-5.2 (753B) on consumer Blackwell GPUs using 2-bit experts and FP4 recovery with the vLLM-Moet patch.
Tag vllm
The Social Emergence of Multi-Agent Collaboration – Optimising Gemma4
A week-long experiment with 100+ agents optimizing Gemma 4 revealed surprising social emergence, self-policing, and a 5x speed improvement in vLLM.