FreeToken runs a 753B model on one GPU

FreeToken, an Apache-2.0 edge-native mixture-of-experts engine, reports serving a 753-billion-parameter GLM-5.2 on one workstation GPU, alongside a 35B model on an 8GB laptop GPU and a 284B model on a gaming desktop. Its open-source repository and the accompanying paper describe CPU-GPU co-execution, expert caching, and runtime memory management rather than a promise that every workstation delivers the same speed. Builders with sensitive workloads should benchmark this local path before buying cloud capacity, while keeping a throughput fallback, extending the archive’s local-agent analysis.