Open Source

This GLM 5.2 Benchmark on 6 RTX 3090s Hit 7.8 Tok/s — But Is It Actually Good?

Community reports real-world inference speeds for the new open-source GLM 5.2 model.

Deep Dive

Reddit user neverbyte shared GLM 5.2 local inference stats using llama.cpp on a system with 6x RTX 3090, 128GB DDR5, and an i7-13700K. With unsloth UD-IQ2_M quantization and 90K context at Q8_0 KV, they got 7.8 tokens/sec generation and roughly 40 tokens/sec prompt processing.

Key Points
  • GLM 5.2 inference reported by neverbyte using llama.cpp on 6x RTX 3090
  • Generation speed: 7.8 tok/s; prompt processing: 40 tok/s at 90K context
  • Quantization: unsloth UD-IQ2_M with Q8_0 KV cache; system: 128GB DDR5, i7-13700K

Why It Matters

Provides a real-world performance baseline for running GLM 5.2 locally on high-end multi-GPU setups.

📬 Get the top 10 AI stories daily