作者： Jakub Rusinowski · 最后更新： 2026年7月21日
Best all-round pick: GPT-oss 120B
The NVIDIA DGX Spark has 128 GB of VRAM, of which about 128 GB is available to a model. The largest model it can hold is Qwen3.8-Flash-Next (180B, 109.5 GB at Q4_K_M).
| Model | Score | Memory at Q4_K_M | Context | Licence |
|---|---|---|---|---|
| GPT-oss 120B | 95.8 | 71.3 GB | 128K | Apache-2.0 |
| Qwen 3.5 122B-A10B (MoE) | 94.2 | 74.5 GB | 125K | Apache 2.0 |
| Qwen 3.5 122B-A10B | 92.5 | 74.5 GB | 128K | Apache-2.0 |
| GLM-5.1 72B | 91.2 | 44.3 GB | 125K | MIT |
| Command R+ (104B) | 91.2 | 63.6 GB | 125K | CC-BY-NC |
| Llama 4 Scout 17B | 91.2 | 66.6 GB | 10M | Llama Community |
| Workload | Recommended on this GPU |
|---|---|
| Coding | Devstral-2 123B (99) Qwen 3.5 122B-A10B (MoE) (96.5) Qwen3-Coder 80B-A3B (MoE) (96.2) |
| General assistant | GPT-oss 120B (95.8) Qwen 3.5 122B-A10B (MoE) (94.2) Qwen 3.5 122B-A10B (92.5) |
| Reasoning | GLM-5.1 72B (98.4) Qwen 3.5 122B-A10B (MoE) (97) Gemma 4 27B ⭐ (95.9) |
| RAG | Llama 4 Scout 17B (93.6) Qwen3.8-Flash-Next (93) Gemma 4 31B (92.8) |
| Agents | Devstral-2 123B (95.2) Qwen3.8-Flash-Next (93.9) Ternary Bonsai 27B (89) |
| Vision | Qwen 3.5 72B (96.9) Llama 3.2 90B Vision Instruct (95.2) Qwen3.8-Flash-Next (93.7) |
GPT-oss 120B is the strongest all-round pick that fits its 128 GB.
Qwen3.8-Flash-Next — 180B parameters, needing 109.5 GB at Q4_K_M.
About 128 GB of its 128 GB, before the desktop and runtime overhead are accounted for.