Model catalogue
Start with a signed recipe tested on DGX Spark, or see what we are testing next. SparkPilot handles the download and setup for recipes that are ready.
1 ready to run3 in testing
NVIDIA · NVFP4
Everyday chat, agents and long documents
A 30B model with 3B active parameters per token, tool calling and presets for long conversations. Download it in SparkPilot with the settings we tested on DGX Spark.
Ready in SparkPilot
Tested on DGX Spark. Signature verified in the app.
Hardware testing and app integration are separate steps; these models are not yet installable in the current TestFlight build.
Qwen · NVFP4
Hard problems and careful answers
A dense 27B model for reasoning, tools and images. The recipe uses about 24 GB for weights, before context and runtime memory.
Qwen · NVFP4
Coding and development
Verified on one DGX Spark at 262,144 tokens, with 30.0 tokens/sec for prose and 45.1 for code in our standalone tests. SparkPilot integration is undergoing security and managed-install validation; this model is not available to install from the app yet.
DeepSeek · EXL3 3.0 bpw
Deep reasoning on one Spark
A text model for reasoning and tool use, with a 384K context in this recipe. Around 100 GB of weights makes memory planning important.
These speeds come from Artificial Analysis testing provider-hosted models. They are not speeds for these quantised recipes on your Spark; SparkPilot measurements, where available, are listed separately in the recipe details.
Scores are Artificial Analysis testing the model as its maker serves it, not this quantised recipe on a Spark. The index (v4.2) combines ten evaluations, including agentic, coding and long-context tasks.
On smaller screens, scroll the table sideways to see every column.
| Model | Provider index | Provider speed | Context | Weights | Recipe status |
|---|---|---|---|---|---|
| 16#139 | 294tokens/sec | 64K to 1M | 24 GB | Validated · signed | |
| 41#34 | 47tokens/sec | 262K native, 1M with YaRN | about 24 GB | In testing | |
| 46#29 | 74tokens/sec | 262,144 tokens verified | 105.94 GB download | Spark verified · app integration pending | |
| 41#44 | 130tokens/sec | 384K | about 100 GB | In testing |
Source: Artificial Analysis Recorded 7 September 2026; rankings and speeds can change.
Weight size is not total memory use. Context, concurrent requests and the inference engine also need memory.
The catalogue is a public, signed file. Your Spark verifies the signature before it changes anything, so the website holds no authority of its own.
SparkPilot