The DGX Spark gets a lot of "is it actually fast?" questions. The honest answer depends on the recipe: which quant, which engine, whether speculative decoding is on, and how many Sparks you link together.
Mia's AI Lab publishes open-source DGX Spark recipes (install scripts, engine work and benchmarks on GitHub) for the newest open models. All of the numbers below are theirs. I collected them, checked each decode figure against its recipe page or README, and put them in one table. Full table with links: getrigbase.com/dgx-spark/.
The short version
- GLM-5.3 Flash (320B total, 18B active) does not fit one Spark at 4-bit, which is about 160 GB of weights. Mia's TensorFold recipe does 60.4 tok/s on 2 Sparks, 77.6 on 3 and 95.2 on 4, single stream, with prefill of 1,979–2,583 tok/s on a 32k prompt.
- Qwen3.8 Flash-Next runs on a single Spark at 48.7 tok/s (NVFP4, vLLM); Mia's TensorFold recipe gives a 46–55 tok/s range.
- DeepSeek V4.1 Flash reaches 74.9 tok/s on 2 Sparks (1017 checkpoint, EXL3 2.9-bit + DSpark). That works because about 196.6B of its parameters are Engram lookup tables that do not have to sit in accelerator memory (vLLM recipe).
- Full GLM-5.3 (753B) needs 3 Sparks at 2.75 bits (≈259 GB of weights): 31.3 tok/s.
All 19 setups
Decode is single-stream tokens/sec. Prefill is prompt processing on a 32k prompt, where it was listed.
| Model | Sparks | Quant | Engine | Decode tok/s | Prefill tok/s | Context | Source |
|---|---|---|---|---|---|---|---|
| Ling 3.0 Flash | 1 | INT4 + DSpark | SGLang | 83 | – | 256k | recipe |
| Muse Glimmer 30B | 1 | NVFP4 + DFlash | vLLM | 35 | – | 256k | recipe |
| Nemotron 3.5 Bolt 74B-A9B | 1 | IQ2_XXS + NVFP4 + LiLiCorr | TensorFold Zig | 71 | 1,375 | 262k · up to 16 at once | recipe |
| Ornith 1.5 122B | 1 | EXL3 4-bit + DFlash2 | TensorFold Zig | 52.6 | 1,955 | 262k · 4 at once | recipe |
| Qwen3.6 35B-A3B | 1 | NVFP4 + MTP | vLLM | 95.1 | – | 262k | recipe |
| Qwen3.8 27B | 1 | MLX 4-bit + DFlash2 | TensorFold | 58.3 | 1,633 | 262k · 1M with YaRN | recipe |
| Qwen3.8 27B | 1 | NVFP4 + DSpark | SGLang | 18.3 | – | 262k · 1M with YaRN | recipe |
| Qwen3.8 Flash-Next | 1 | NVFP4 | vLLM | 48.7 | 1,769 | 262k · 512k with YaRN | recipe |
| Qwen3.8 Flash-Next | 1 | FP4 QAD or INT4-AR + MTP | TensorFold Zig | 46-55 | 2,430 | 262k · up to 5 at once | recipe |
| DeepSeek V4.1 Flash | 2 | 1017 checkpoint, EXL3 2.9-bit + DSpark | TensorFold Zig | 74.9 | 2,175 | 1M | recipe |
| DeepSeek V4.1 Flash | 2 | EXL3 2.9-bit | vLLM | 31.6 | – | 600k | recipe |
| GLM-5.3 Flash | 2 | EXL3 4-bit + DFlash2 | TensorFold | 60.4 | 1,978.9 | 1M | recipe |
| MiMo V2.6 Flash | 2 | MXFP4 | SGLang | 35 | – | 1M | recipe |
| Qwen3.8 Flash-Next | 2 | NVFP4 or INT4-AR + MTP | TensorFold Zig | 67.4 | 2,575 | 1M · ×3 at full context | recipe |
| Qwen3.8 Flash-Next | 2 | NVFP4 | vLLM | 54.4 | 2,962 | 262k · 1M with YaRN | recipe |
| GLM-5.3 | 3 | EXL3 2.75-bit + DSpark | TensorFold | 31.3 | 591.2 | ~500k | recipe |
| GLM-5.3 Flash | 3 | EXL3 4-bit + DFlash2 | TensorFold | 77.6 | 2,064.2 | 1M | recipe |
| DeepSeek V4.1 Flash | 4 | 1017 checkpoint, FP8 + MXFP4 + DSpark | TensorFold Zig | 65.8 | 4,127.2 | 1M · up to 16 at once | recipe |
| GLM-5.3 Flash | 4 | EXL3 4-bit + DFlash2 | TensorFold | 95.2 | 2,583.3 | 1M | recipe |
Most rows use speculative decoding (MTP, DFlash, DSpark, LiLiCorr). It helps predictable text such as code far more than free prose, so do not compare these 1:1 with plain-decode numbers from other GPUs.
Memory math in one line
Each Spark has 128 GB of unified memory (NVIDIA). Weights take about parameters × bits ÷ 8. NVFP4 is about 4.5 bits per weight because every 16 values share an FP8 scale (NVIDIA). Leave room for the KV cache on top. So:
- 320B at 4 bits is about 160 GB: 2 Sparks.
- 753B at 2.75 bits is about 259 GB: 3 Sparks.
- A 27–35B model at 4 bits is 14–20 GB: one Spark with lots of room for context.
What I left out
Figures I could not find again in the recipe text are not in the table. That covers some card numbers that differ from the README, a chart-only number, and third-party numbers for other GPUs. Setups that are "not benchmarked yet" are also out. The full list is in the import script's skip log.
The table updates as Mia publishes new recipes: getrigbase.com/dgx-spark/. Comparing against a Mac Studio or Strix Halo? See DGX Spark vs Mac Studio vs Strix Halo.
Numbers: Mia's AI Lab (mia-ai.net), read 2026-10-27. RigBase did not re-run them.
Top comments (0)