Capability rankings for DGX Spark owners
The best open models that fit on your Sparks
Ranked on Cybench, CyberGym, LiveCodeBench, SWE-bench Pro, Terminal-Bench. Each model is shown with the recipe your cluster can run, and a score only counts if it was measured on that exact model version and build.
How many Sparks do you have?
More Sparks can unlock other recipes, and run the same recipe faster
Best picks for 1 Spark
Every pick is on the frontier: nothing that fits is both faster and more capable
No picks yet: not enough scores to draw the frontier
None of the 11 recipes that fit 1 Spark has scores that count on 3 of the 5 benchmarks it needs to be ranked overall, so none can be placed on the chart. Each one appears here at its measured decode speed once it does.
10 of the 11 already have a decode speed measured on 1 Spark; what’s missing is scores on their exact builds.
Submit a scoreLeaderboard
9 models · 0 of 45 scores count for these recipes
| # | Model | Recipe for 1 Spark |
Counted | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Not ranked: fewer than 3 of 5 benchmarks count for this recipe | |||||||||
| – | Laguna S 2.1 poolside · 118B · 8B active |
poolside NVFP4poolside1 Spark · 71 GB77 t/s | 0/5 · +2 | — | — | — | 59.4@BF16 | 70.2@BF16 | |
| – | Ling-3.0-flash inclusionAI · 124B · 5.1B active |
inclusionAI INT4inclusionAI1 Spark83 t/s | 0/5 · +3 | — | — | 84.0●@BF16 | 56.6@BF16 | 50.2●@BF16 | |
| – | Muse Glimmer 30B Meta Superintelligence Lab · 29.6B · 29.6B active |
Red Hat NVFP4Red Hat AI1 Spark · 19 GB35 t/s | 0/5 · +2 | — | — | — | 51.2@BF16 | 51.7●@BF16 | |
| – | Nemotron 3.5 Bolt 74B-A9B NVIDIA · 74B · 9B active |
NVIDIA 24GB packNVIDIA1 Spark · 24 GB71 t/s | 0/5 · +1 | — | — | — | — | 51.02.1 | |
| – | Nemotron 3.5 Lightning 30B-A3B NVIDIA · 30B · 3B active |
NVIDIA NVFP4NVIDIA1 Spark | 0/5 · +1 | — | — | — | — | 23.52.1 | |
| – | Ornith-1.5-122B Ornith AI · 125.09B |
Mia's EXL3 EAQ 4-bitMia's AI Lab1 Spark · 67 GB53 t/s | 0/5 · +3 | — | 80.0@BF16 | — | 76.9@BF16 | 81.2@BF16 | |
| – | Qwen3.6-35B-A3B Qwen · 35B · 3B active |
Unsloth NVFP4Unsloth1 Spark95 t/s | 0/5 · +3 | — | — | 80.4@BF16 | 49.5@BF16 | 0.0●@BF16 | |
| – | Qwen3.8-27B Qwen · 27.8B · 27.8B active |
Vontra MLX 4-bitVontra1 Spark · 19 GB58 t/s · +1 more | 0/5 · +3 | — | — | 90.3@BF16 | 61.7@BF16 | 69.4●@NVFP4 | |
| – | Qwen3.8-Flash-Next Qwen · 180B · 6B active |
NVIDIA FP4-LocalNVIDIA1 Spark · 103 GB56 t/s · +1 more | 0/5 · +3 | — | — | 95.0●@NVFP4 | 62.9●@BF16 | 86.4●@NVFP4 | |
Index: the mean of a recipe’s percentiles, among the models that fit, across the benchmarks that count for it. A recipe needs 3 of 5 to be ranked. Dashed cells were measured on a different build, model version, test version or a subset, and are left out. Hover a score for its source; click a column to sort by it.
How complete the grid is
Of 45 possible scores for the 9 models that fit 1 Spark, 0 were measured on the recipe’s exact model and build. Another 21 exist but were measured on something else.
Scores we need next
Benchmarks that don’t yet count for the top recipes. A run on the recipe’s exact build, on a Spark, is the most useful contribution.
- Laguna S 2.1 × CybenchNot reported for this modelSubmit a score
- Laguna S 2.1 × CyberGymNot reported for this modelSubmit a score
- Laguna S 2.1 × LiveCodeBenchNot reported for this modelSubmit a score
- Laguna S 2.1 × SWE-bench ProOnly measured on poolside BF16; this recipe runs poolside NVFP4Submit a score
- Laguna S 2.1 × Terminal-BenchOnly measured on poolside BF16; this recipe runs poolside NVFP4Submit a score
- Ling-3.0-flash × CybenchNot reported for this modelSubmit a score
- Ling-3.0-flash × CyberGymNot reported for this modelSubmit a score
- Ling-3.0-flash × LiveCodeBenchOnly measured on inclusionAI BF16; this recipe runs inclusionAI INT4Submit a score