Which local LLMs can the NVIDIA DGX Spark run?
Unified memory — the GPU can use what Linux and the desktop leave free
Assumes 8k context and an f16 KV cache. Verdicts are shown for each memory size.
What to expect from the NVIDIA DGX Spark
The DGX Spark is NVIDIA's desktop built on its GB10 chip. NVIDIA sells it with 128GB, and its partner makers sell the same chip under their own names, such as the ASUS Ascent GX10, the Dell Pro Max with GB10 and the Lenovo ThinkStation PGX. NVIDIA lists six partners for a 64GB configuration: Acer, ASUS, Dell, GIGABYTE, HP and MSI. Both sizes have 273 GB/s of bandwidth, so a model that fits gets the same speed estimate on either. CanRun counts what Linux and the desktop leave free for the GPU: 116.9 GB on the 128GB model, based on a public boot log, and 48.2 GB on the 64GB model, worked out from the same log because none has been published for that size.
That share is where the two sizes split. With 64GB, gpt-oss-20b and Qwen3.6 35B-A3B run great, Llama 3.3 70B fits at Q4_K_M but runs slowly, and gpt-oss-120b does not load. With 128GB, gpt-oss-120b runs great and Qwen3.5 122B-A10B runs well, both backed by public benchmarks.
NVIDIA supports linking up to four units over the built-in ConnectX-7 networking: two or three directly by cable, and for four, NVIDIA recommends a switch. The tables below also show two and four linked units running llama.cpp, as theoretical estimates.
Specs
- Memory options
- 64 GB · 128 GB
- Memory bandwidth
- 273.0 GB/s
- FP16 compute
- 62.5 TFLOPS
- Launch year
- 2025
- Price
- $3,999 launch MSRP (128 GB)
- Also written as
- DGX Spark, GB10, ASUS Ascent GX10, Dell Pro Max with GB10, HP ZGX Nano G1n, Lenovo ThinkStation PGX, MSI EdgeXpert, Acer Veriton GN100, GIGABYTE AI TOP ATOM
Verdicts at a glance
At Q4_K_M (or the closest available quant) with 8k context.
- 64 GBRuns great31 modelsRuns well9 modelsRuns slowly10 modelsWon't run9 models3 more models run at a lower quant.
- 128 GBRuns great34 modelsRuns well12 modelsRuns slowly10 modelsWon't run3 models1 more model runs at a lower quant.
- 64 GB × 2Runs great3 modelsRuns well3 modelsRuns slowly0 modelsWon't run3 modelsOne unit is enough50 models
- 128 GB × 2Runs great0 modelsRuns well3 modelsRuns slowly0 modelsWon't run0 modelsOne unit is enough56 models
- 64 GB × 4Runs great0 modelsRuns well8 modelsRuns slowly0 modelsWon't run1 model1 more model runs at a lower quant.One unit is enough50 models
- 128 GB × 4Runs great0 modelsRuns well2 modelsRuns slowly1 modelWon't run0 modelsOne unit is enough56 models
Rows marked “× 2” or “× 4” link two or four units over the built-in ConnectX-7 networking and run llama.cpp RPC, which splits the layers across the units. Memory adds up, but each token still passes through every unit, and every hop between units adds time to each token. So linking is never faster than CanRun’s one-unit estimate: in public runs, gpt-oss-120b ran about 20% slower on two units than on one. A model whose baseline quant already fits fully on the GPU of one unit with the same memory therefore shows “One unit is enough” on these rows instead of a verdict and speed. The verdicts on these rows are theoretical estimates (±30%). vLLM tensor parallelism, which NVIDIA’s official playbooks use, can be faster. Three units can be checked in the calculator. On the “× 2” and “× 4” rows of this summary, the verdict counts include only the models that need linking, and “One unit is enough” counts the rest.
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
Every model on the NVIDIA DGX Spark
Scroll sideways to see every column.
| Model | Verdict | Speed | Quant | Memory | Runs as | Context | Notes |
|---|---|---|---|---|---|---|---|
| MiniCPM5 1B | Runs great | est. 238.9 tok/s191.1–286.7calibrated estimate ±20% | Q4_K_M | 1.5 / 116.9 GB | Unified memory | up to 128k |
|
| Llama 3.2 1B | Runs great | est. 197.5 tok/s158.0–236.9calibrated estimate ±20% | Q4_K_M | 1.7 / 116.9 GB | Unified memory | up to 128k | — |
| Ling-3.0 Tiny | Runs great | est. 179.8 tok/s125.8–233.7theoretical estimate ±30% | Q4_K_M | 5.5 / 116.9 GB | Unified memory | up to 128k |
|
| EXAONE 4.0 1.2B | Runs great | est. 162.1 tok/s129.7–194.6calibrated estimate ±20% | Q4_K_M | 1.9 / 116.9 GB | Unified memory | up to 64k |
|
| LFM2.5 2.6B | Runs great | est. 118.0 tok/s82.6–153.4theoretical estimate ±30% | Q4_K_M | 2.4 / 116.9 GB | Unified memory | up to 128k |
|
| HyperCLOVA X SEED 1.5B | Runs great | est. 117.3 tok/s93.8–140.8calibrated estimate ±20% | Q4_K_M | 2.4 / 116.9 GB | Unified memory | up to 16k | — |
| Gemma 4 E2B | Runs great | est. 115.9 tok/s81.1–150.6theoretical estimate ±30% | Q4_K_M | 3.7 / 116.9 GB | Unified memory | up to 128k |
|
| MiniCPM5 2B | Runs great | est. 111.4 tok/s89.1–133.6calibrated estimate ±20% | Q4_K_M | 2.5 / 116.9 GB | Unified memory | up to 64k |
|
| Mellum2.1 12B-A2.5B Thinking | Runs great | est. 104.4 tok/s83.5–125.2calibrated estimate ±20% | Q4_K_M | 8.8 / 116.9 GB | Unified memory | up to 128k |
|
| MiniCPM-V 4.7 35B-A3B | Runs great | est. 76.0 tok/s60.8–91.2calibrated estimate ±20% | Q4_K_M | 21.9 / 116.9 GB | Unified memory | up to 128k |
|
| Spark-X2.5 4B | Runs great | est. 73.4 tok/s58.7–88.1calibrated estimate ±20% | Q4_K_M | 3.5 / 116.9 GB | Unified memory | up to 64k |
|
| Llama 3.2 3B | Runs great | est. 72.0 tok/s57.6–86.3calibrated estimate ±20% | Q4_K_M | 3.5 / 116.9 GB | Unified memory | up to 64k | — |
| gpt-oss-20b | Runs great | est. 71.7 tok/s57.4–86.0calibrated estimate ±20% | MXFP4 | 12.9 / 116.9 GB | Unified memory | up to 128k |
|
| Qwen3.6 35B-A3B | Runs great | est. 70.5 tok/s56.4–84.6calibrated estimate ±20% | Q4_K_M | 23.0 / 116.9 GB | Unified memory | up to 128k |
|
| Nemotron 3.5 Lightning 30B-A3B | Runs great | est. 69.8 tok/s48.8–90.7theoretical estimate ±30% | Q4_K_M | 26.1 / 116.9 GB | Unified memory | up to 256k |
|
| Qwen3.5 4B | Runs great | est. 68.7 tok/s55.0–82.5calibrated estimate ±20% | Q4_K_M | 3.7 / 116.9 GB | Unified memory | up to 64k |
|
| Kanana 1.5 15.7B-A3B | Runs great | est. 62.2 tok/s49.7–74.6calibrated estimate ±20% | Q4_K_M | 12.1 / 116.9 GB | Unified memory | up to 32k | — |
| Qwen3.5 35B-A3B | Runs great | est. 61.9 tok/s49.5–74.2calibrated estimate ±20% | Q4_K_M | 23.4 / 116.9 GB | Unified memory | up to 128k |
|
| Gemma 4 E4B | Runs great | est. 61.5 tok/s43.1–80.0theoretical estimate ±30% | Q4_K_M | 5.7 / 116.9 GB | Unified memory | up to 128k |
|
| Qwen3 30B-A3B (2507) | Runs great | est. 60.7 tok/s48.6–72.9calibrated estimate ±20% | Q4_K_M | 19.9 / 116.9 GB | Unified memory | up to 64k | — |
| gpt-oss-120b | Runs great | 58.7 tok/s8k estimate 50.8 tok/s (40.6–60.9, calibrated estimate ±20%)measured (1 run, 2k context) | MXFP4 | 64.3 / 116.9 GB | Unified memory | up to 64k |
|
| Ling-3.0 Flash | Runs great | est. 52.8 tok/s36.9–68.6theoretical estimate ±30% | Q4_K_M | 77.7 / 116.9 GB | Unified memory | up to 256k |
|
| Kolibri 1 | Runs great | est. 52.2 tok/s41.7–62.6calibrated estimate ±20% | Q4_K_M | 50.5 / 116.9 GB | Unified memory | up to 128k |
|
| Gemma 4 26B-A4B | Runs great | 47.0 tok/s8k estimate 51.5 tok/s (41.2–61.8, calibrated estimate ±20%)measured (1 run, 32k context) | Q4_K_M | 17.7 / 116.9 GB | Unified memory | up to 128k |
|
| Qwen3 8B | Runs great | 43.9 tok/s8k estimate 34.1 tok/s (27.3–41.0, calibrated estimate ±20%)measured (1 run, 512 context) | Q4_K_M | 6.8 / 116.9 GB | Unified memory | up to 8k |
|
| K2-Horizon MoVA 36B-A4B | Runs great | est. 37.7 tok/s26.4–49.1theoretical estimate ±30% | Q4_K_M | 24.6 / 116.9 GB | Unified memory | up to 8k |
|
| DeepSeek R1 Distill Llama 8B | Runs great | est. 35.5 tok/s28.4–42.6calibrated estimate ±20% | Q4_K_M | 6.6 / 116.9 GB | Unified memory | up to 16k |
|
| Kanana 1.5 8B | Runs great | est. 35.5 tok/s28.4–42.6calibrated estimate ±20% | Q4_K_M | 6.6 / 116.9 GB | Unified memory | up to 32k | — |
| Llama 3.1 8B | Runs great | est. 35.5 tok/s28.4–42.6calibrated estimate ±20% | Q4_K_M | 6.6 / 116.9 GB | Unified memory | up to 32k | — |
| Qwen3.5 9B | Runs great | 35.2 tok/s8k estimate 34.7 tok/s (27.8–41.6, calibrated estimate ±20%)measured (1 run, 4k context) | Q4_K_M | 6.7 / 116.9 GB | Unified memory | up to 32k |
|
| Nanbeige4.2 3B | Runs great | est. 34.8 tok/s24.3–45.2theoretical estimate ±30% | Q4_K_M | 4.7 / 116.9 GB | Unified memory | up to 8k |
|
| DeepSeek R1 0528 Qwen3 8B | Runs great | est. 34.1 tok/s27.3–41.0calibrated estimate ±20% | Q4_K_M | 6.8 / 116.9 GB | Unified memory | up to 8k |
|
| K2-Horizon 7B | Runs great | est. 31.3 tok/s25.1–37.6calibrated estimate ±20% | Q4_K_M | 7.4 / 116.9 GB | Unified memory | up to 8k |
|
| Gemma 3 12B | Runs great | est. 27.2 tok/s21.7–32.6calibrated estimate ±20% | Q4_K_M | 8.4 / 116.9 GB | Unified memory | up to 32k | — |
| Gemma 4 12B | Runs well | est. 29.4 tok/s23.5–35.2calibrated estimate ±20% | Q4_K_M | 7.8 / 116.9 GB | Unified memory | up to 256k |
|
| Qwen3.5 122B-A10B | Runs well | 29.0 tok/s8k estimate 24.0 tok/s (19.2–28.8, calibrated estimate ±20%)measured (1 run, 4k context) | Q4_K_M | 79.0 / 116.9 GB | Unified memory | up to 256k |
|
| Laguna S 2.1 | Runs well | est. 28.4 tok/s22.7–34.1calibrated estimate ±20% | Q4_K_M | 68.6 / 116.9 GB | Unified memory | up to 128k |
|
| Solar Open 100B | Runs well | est. 22.8 tok/s18.3–27.4calibrated estimate ±20% | Q4_K_M | 64.4 / 116.9 GB | Unified memory | up to 32k |
|
| HyperCLOVA X SEED Think 14B | Runs well | est. 20.9 tok/s14.6–27.2theoretical estimate ±30% | Q4_K_M | 10.8 / 116.9 GB | Unified memory | up to 32k |
|
| Qwen3 14B | Runs well | est. 20.6 tok/s16.5–24.7calibrated estimate ±20% | Q4_K_M | 10.9 / 116.9 GB | Unified memory | up to 32k |
|
| DeepSeek R1 Distill Qwen 14B | Runs well | est. 20.1 tok/s16.1–24.1calibrated estimate ±20% | Q4_K_M | 11.2 / 116.9 GB | Unified memory | up to 32k |
|
| Phi-4 | Runs well | est. 19.8 tok/s15.9–23.8calibrated estimate ±20% | Q4_K_M | 11.3 / 116.9 GB | Unified memory | up to 16k | — |
| Mistral Small 3.2 24B | Runs well | est. 13.6 tok/s10.9–16.3calibrated estimate ±20% | Q4_K_M | 16.2 / 116.9 GB | Unified memory | up to 64k |
|
| Gemma 3 27B | Runs well | est. 12.4 tok/s9.9–14.8calibrated estimate ±20% | Q4_K_M | 17.8 / 116.9 GB | Unified memory | up to 64k | — |
| Qwen3.6 27B | Runs well | est. 12.3 tok/s9.8–14.7calibrated estimate ±20% | Q4_K_M | 17.9 / 116.9 GB | Unified memory | up to 8k |
|
| Muse Glimmer 30B | Runs well | est. 12.2 tok/s9.8–14.7calibrated estimate ±20% | Q4_K_M | 18.0 / 116.9 GB | Unified memory | up to 16k |
|
| Qwen3.8 27B | Runs slowly | est. 11.8 tok/s9.5–14.2calibrated estimate ±20% | Q4_K_M | 18.6 / 116.9 GB | Unified memory | up to 256k |
|
| Qwen3.5 27B | Runs slowly | 11.3 tok/s8k estimate 12.1 tok/s (9.7–14.5, calibrated estimate ±20%)measured (1 run, 4k context) | Q4_K_M | 18.2 / 116.9 GB | Unified memory | up to 8k |
|
| Granite 4.2 30B | Runs slowly | est. 10.7 tok/s8.6–12.9calibrated estimate ±20% | Q4_K_M | 20.4 / 116.9 GB | Unified memory | up to 128k |
|
| EXAONE 4.0 32B | Runs slowly | est. 10.7 tok/s8.6–12.9calibrated estimate ±20% | Q4_K_M | 20.5 / 116.9 GB | Unified memory | up to 128k |
|
| Qwen3 32B | Runs slowly | 10.7 tok/s8k estimate 9.7 tok/s (7.8–11.7, calibrated estimate ±20%)measured (2 runs, 512 context) | Q4_K_M | 22.5 / 116.9 GB | Unified memory | up to 32k |
|
| EXAONE 4.5 33B | Runs slowly | est. 10.3 tok/s8.3–12.4calibrated estimate ±20% | Q4_K_M | 21.2 / 116.9 GB | Unified memory | up to 256k |
|
| Gemma 4 31B | Runs slowly | 10.1 tok/s8k estimate 11.2 tok/s (9.0–13.5, calibrated estimate ±20%)measured (1 run, 512 context) | Q4_K_M | 19.6 / 116.9 GB | Unified memory | up to 256k |
|
| DeepSeek R1 Distill Qwen 32B | Runs slowly | est. 9.7 tok/s7.7–11.6calibrated estimate ±20% | Q4_K_M | 22.6 / 116.9 GB | Unified memory | up to 128k |
|
| Nanbeige4.3 33B | Runs slowly | est. 9.4 tok/s7.6–11.3calibrated estimate ±20% | Q4_K_M | 23.1 / 116.9 GB | Unified memory | up to 128k |
|
| Llama 3.3 70B | Runs slowly | 4.7 tok/s8k estimate 4.7 tok/s (3.8–5.7, calibrated estimate ±20%)measured (2 runs, 512 context) | Q4_K_M | 45.8 / 116.9 GB | Unified memory | up to 128k | — |
| Solar Open 2 250B | Won't runTry Q2_K (heavy quality loss): Runs well | est. 22.0 tok/s17.6–26.3calibrated estimate ±20% | Q2_K | 96.4 / 116.9 GB | Unified memory | up to 128k |
|
| DeepSeek V4 Flash 0731 | Won't run | — | MXFP4 | needs 155.6 GB | — | — |
|
| GLM-5.3 Flash | Won't run | — | Q4_K_M | needs 182.1 GB | — | — |
|
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Verdict by memory size
| Model | 64 GB | 128 GB | 64 GB × 2 | 128 GB × 2 | 64 GB × 4 | 128 GB × 4 |
|---|---|---|---|---|---|---|
| MiniCPM5 1B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Llama 3.2 1B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Ling-3.0 Tiny | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| EXAONE 4.0 1.2B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| LFM2.5 2.6B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| HyperCLOVA X SEED 1.5B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Gemma 4 E2B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| MiniCPM5 2B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Mellum2.1 12B-A2.5B Thinking | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| MiniCPM-V 4.7 35B-A3B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Spark-X2.5 4B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Llama 3.2 3B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| gpt-oss-20b | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.6 35B-A3B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Nemotron 3.5 Lightning 30B-A3B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.5 4B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Kanana 1.5 15.7B-A3B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.5 35B-A3B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Gemma 4 E4B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3 30B-A3B (2507) | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| gpt-oss-120b | Won't run | Runs great | Runs great | One unit is enough | Runs well | One unit is enough |
| Ling-3.0 Flash | Won't run | Runs great | Runs great | One unit is enough | Runs well | One unit is enough |
| Kolibri 1 | Won't run | Runs great | Runs great | One unit is enough | Runs well | One unit is enough |
| Gemma 4 26B-A4B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3 8B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| K2-Horizon MoVA 36B-A4B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| DeepSeek R1 Distill Llama 8B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Kanana 1.5 8B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Llama 3.1 8B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.5 9B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Nanbeige4.2 3B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| DeepSeek R1 0528 Qwen3 8B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| K2-Horizon 7B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Gemma 3 12B | Runs great | Runs great | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Gemma 4 12B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.5 122B-A10B | Won't run | Runs well | Runs well | One unit is enough | Runs well | One unit is enough |
| Laguna S 2.1 | Won't run | Runs well | Runs well | One unit is enough | Runs well | One unit is enough |
| Solar Open 100B | Won't run | Runs well | Runs well | One unit is enough | Runs well | One unit is enough |
| HyperCLOVA X SEED Think 14B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3 14B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| DeepSeek R1 Distill Qwen 14B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Phi-4 | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Mistral Small 3.2 24B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Gemma 3 27B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.6 27B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Muse Glimmer 30B | Runs well | Runs well | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.8 27B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3.5 27B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Granite 4.2 30B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| EXAONE 4.0 32B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Qwen3 32B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| EXAONE 4.5 33B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Gemma 4 31B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| DeepSeek R1 Distill Qwen 32B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Nanbeige4.3 33B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Llama 3.3 70B | Runs slowly | Runs slowly | One unit is enough | One unit is enough | One unit is enough | One unit is enough |
| Solar Open 2 250B | Won't run | Won't run | Won't run | Runs well | Runs well | Runs well |
| DeepSeek V4 Flash 0731 | Won't run | Won't run | Won't run | Runs well | Runs well | Runs well |
| GLM-5.3 Flash | Won't run | Won't run | Won't run | Runs well | Won't run | Runs slowly |
Measured results on the NVIDIA DGX Spark
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Model | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| gpt-oss-120b | MXFP4 | llama.cpp | 2k | 2,444.0 | 58.7 | CUDA, no mmap | github.com | 2026-02-05 |
| Gemma 4 31B | Q4_K_M | llama.cpp | 512 | 687.0 | 10.1 | CUDA | github.com | 2026-04-14 |
| Gemma 4 26B-A4B | Q4_K_M | llama.cpp | 32k | — | 47.0 | CUDA, depth 32768 | github.com | 2026-06-11 |
| Qwen3.5 35B-A3B | Q8_0 | llama.cpp | 2k | 2,572.0 | 49.2 | CUDA | github.com | 2026-03-12 |
| Qwen3.5 9B | Q4_K_M | llama.cpp | 4k | — | 35.2 | — | github.com | 2026-08-13 |
| Qwen3.5 27B | Q4_K_M | llama.cpp | 4k | — | 11.3 | — | github.com | 2026-08-13 |
| Qwen3.5 122B-A10B | Q4_K_M | llama.cpp | 4k | — | 29.0 | — | github.com | 2026-08-13 |
| Llama 3.3 70B | Q4_K_M | llama.cpp | 512 | 284.0 | 4.7 | CUDA, Vulkan | github.com | 2025-11-22 |
| Llama 3.3 70B | Q4_K_M | llama.cpp | 512 | — | 4.6 | CUDA | github.com | 2026-08-12 |
| Llama 3.3 70B | Q8_0 | llama.cpp | 512 | — | 2.9 | CUDA | github.com | 2026-08-12 |
| Llama 3.3 70B | Q2_K | llama.cpp | 512 | — | 7.2 | CUDA | github.com | 2026-08-12 |
| Qwen3 32B | Q4_K_M | llama.cpp | 512 | 630.0 | 10.5 | CUDA, Vulkan | github.com | 2025-11-22 |
| Qwen3 32B | Q4_K_M | llama.cpp | 512 | 762.0 | 10.8 | CUDA | github.com | 2025-12-05 |
| Qwen3 8B | Q4_K_M | llama.cpp | 512 | 3,168.0 | 43.9 | CUDA | github.com | 2025-12-05 |
Frequently asked questions
What is the largest model that runs entirely on the NVIDIA DGX Spark?
Ling-3.0 Flash at Q4_K_M (a 77.01 GB file) fits entirely in 128 GB of unified memory with 8k context, at est. 52.8 tok/s (36.9–68.6, theoretical estimate ±30%).
How many local LLMs run well on the NVIDIA DGX Spark?
At Q4_K_M with 8k context, with 128 GB of memory, out of 59 tracked models: 34 run great, 12 run well, 10 run slowly and 3 won't run.
Can the NVIDIA DGX Spark run a 70B model like Llama 3.3 70B?
Runs slowly — Unified memory, Q4_K_M: 4.7 tok/s measured at 512 context (2 runs; 8k estimate 4.7 tok/s, 3.8–5.7, calibrated estimate ±20%).
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.