GLM-5.2 4-bit on 4× DGX Spark achieves 70.8% on Terminal-Bench 2.1

A user quantized the GLM-5.2 (753B MoE) model to Int4-Int8Mix with NVFP4 4-bit KV cache and ran it on four DGX Spark (GB10) systems, achieving a 70.8% score on Terminal-Bench 2.1.

• The setup used TP=4 across 4× DGX Spark nodes with ~273 GB/s bandwidth and a 100K context window enabled by the 4-bit KV cache.
• Inference utilized MTP speculative decode and FULL CUDA graphs, yielding approximately 27.5 tokens per second.
• The result of 63/89 tasks is compared against the official full-precision score of 81.0% on the same Terminus-2 agent scaffold.
• The ~10-point gap includes effects from quantization, a reduced 100K context cap versus the official 256K, and a smaller token budget.

The run demonstrates that a 4-bit open-weight model on consumer-grade hardware can retain approximately 87% of the official full-precision performance on complex agentic benchmarks.