Can DGX Spark 128G run Recursive Flint 40B-A2.5B?
Comfortable: GGUF Q8_0 needs 44.5 GB at 4K, batch 1; headroom 83.5 GB.
Text input, FP16 KV cache, one card. Unified-memory capacity uses the GPU guide’s usable limit; negative headroom may require host offload. Speed source settings may differ from this 4K estimate.
Measured runs
No measured run for this pair. Sizing is an estimate, not a runtime test.
Quant memory at 4K
| Quant | Total GB | Verdict | Weight sizing |
|---|---|---|---|
| GGUF Q3_K_M | 20.6 | Comfortable | Verified file |
| GGUF Q4_K_M | 26.7 | Comfortable | Verified file |
| GGUF Q8_0 | 44.5 | Comfortable | Verified file |
Context capacity · GGUF Q8_0
| Context | Total GB | Verdict |
|---|---|---|
| 4K | 44.5 | Comfortable |
| 16K | 44.8 | Comfortable |
| 32K | 45.3 | Comfortable |
| 128K | 48.2 | Comfortable |
Generation speed
GGUF Q8_0: 108 tok/s · Bandwidth ceiling (not measured)