GLM-5.1 Pushes Mi50 Rigs To Edge

The headline question is not whether GLM-5.1 can be forced onto older AMD cards, but whether the result is worth the pain. The community consensus in the thread is basically “maybe with extreme quantization, but expect ugly tradeoffs.”

• –Official docs pitch GLM-5.1 as a long-horizon coding model with 200K context, tool use, MCP support, and strong agentic workflows, which naturally raises memory and activation pressure.
• –Reddit replies are blunt: 8x32GB = 256GB still looks too tight unless you drop to very aggressive Q2-style quantization, and even then quality is heavily compromised.
• –One commenter flags a key MI50 weakness for this use case: no Infinity Fabric/P2P, which hurts multi-GPU scaling and makes prompt processing the bottleneck.
• –The practical read is that MI50s may be viable for experimentation or heavily offloaded hybrid inference, but they are a poor buy if the goal is smooth local GLM-5.1 performance.
• –If you want a comparable local workload, commenters are already steering toward smaller models like Qwen 397B or Minimax 2.7 instead of buying more aging GPUs.