Can GLM 4.6V run on the DGX Spark?
Yes. At 8K context the best fit is Q6_K, using about 91.4 GB of the 109 GB usable on the DGX Spark. Estimated speed ceiling: ~21 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | 42.3 GB Fits | 43.0 GB Fits | 47.3 GB Fits | 64.5 GB Fits | 47 |
| 3-bit (Q3_K_M) • | 53.0 GB Fits | 53.8 GB Fits | 58.1 GB Fits | 75.3 GB Fits | 37 |
| 4-bit (Q4_K_M) • | 67.1 GB Fits | 67.8 GB Fits | 72.2 GB Fits | 89.4 GB Fits | 29 |
| 5-bit (Q5_K_M) • | 76.6 GB Fits | 77.3 GB Fits | 81.6 GB Fits | 98.9 GB Tight | 25 |
| 6-bit (Q6_K) • | 90.7 GB Fits | 91.4 GB Fits | 95.7 GB Fits | 113 GB No | 21 |
| 8-bit (Q8_0) • | 107 GB Tight | 108 GB Tight | 112 GB No | 129 GB No | 18 |
| 16-bit (F16/BF16) • | 200 GB No | 201 GB No | 206 GB No | 223 GB No | — |
• real GGUF file size from unsloth/GLM-4.6V-GGUF. Speed is a bandwidth ceiling (273 GB/s), not a benchmark. Methodology.
More
GLM 4.6V on every GPU and Mac · Every model that runs on the DGX Spark · Other models that fit 64 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (zai-org/GLM-4.6V), NVIDIA specs. Data as of 2026-11-17.