exo users size up GLM hardware

This is the right instinct, but the budget math is harsher than the enthusiasm: Exo can pool heterogeneous devices, yet GLM-4.7-Flash is a 30B-A3B MoE model, so throughput still depends on real VRAM and interconnect quality.

• –Exo’s appeal is aggregation: it can split work across Macs, GPUs, and CPUs, so a 24GB Mac mini can contribute instead of sitting idle.
• –The catch is that local speed comes from memory headroom, not just model loading; a single 24GB machine is a starter node, not a serious coding-agent box.
• –For a genuinely fast setup, you want either a high-VRAM NVIDIA GPU rig or multiple Apple Silicon boxes linked tightly enough that bandwidth does not erase the gains.
• –If the goal is Claude Code-like iteration speed, a smaller quantized model or hosted GLM plan will usually beat a hobby cluster on simplicity.