NVIDIA DGX Spark vs Framework Desktop 128 GB
DGX Spark · 128 GB unified and Framework Desktop · Ryzen AI Max+ 395 · 128 GB side by side for local AI.
Facts from both hardware pages
| Specification | NVIDIA DGX Spark | Framework Desktop 128 GB |
|---|---|---|
| Memory | 128 GB unified memoryShared by CPU and GPU · source | 128 GB unified memoryShared by CPU and GPU · source |
| Memory bandwidth | 273 GB/sManufacturer spec · not model speed · source | 256 GB/sManufacturer spec · not model speed · source |
| Text · run well | 849+1 tight fit · within 128 GB | 849+1 tight fit · within 128 GB |
| Vision · run well | 1,103within 128 GB | 1,103within 128 GB |
| Image · within memory | 174within 128 GB · documented paths and minimums | 174within 128 GB · documented paths and minimums |
| Video · within memory | 132within 128 GB · documented paths and minimums | 132within 128 GB · documented paths and minimums |
| Audio · within memory | 389within 128 GB · documented paths and minimums | 389within 128 GB · documented paths and minimums |
| Embeddings · within memory | 654within 128 GB · 8-bit estimate | 654within 128 GB · 8-bit estimate |
| Published speed runs | 1413 on this machine · 1 naming only the chip | 44 on this machine |
| Launch price (MSRP) | USD 3,999US · dated 2025-10-15 · not a tracked offer · source | USD 1,999US · dated 2025-02-25 · not a tracked offer · source |
| Where to buy | NVIDIA Marketplacethis configuration · US · recorded 2026-10-18 · store link | No manufacturer store link on file |
Counts use the same memory estimate as each hardware page: text and vision run well within the fast memory (tight fits counted apart); image, video and audio use documented paths and publisher minimums. Formulas
Top text models each runs well
The first 5 rows of each hardware page's Text list, in its order: LiveBench score, then reviewed models, then Hub downloads. Memory is the planning estimate on that machine.
| Rank | NVIDIA DGX Spark | Framework Desktop 128 GB |
|---|---|---|
| #1 | Smaug MiniAbacus AI · LiveBench 76.9 · ~29 GB | Smaug MiniAbacus AI · LiveBench 76.9 · ~29 GB |
| #2 | Qwen3.8 27BAlibaba · LiveBench 75.3 · ~29 GB | Qwen3.8 27BAlibaba · LiveBench 75.3 · ~29 GB |
| #3 | Qwen3.6 27BAlibaba · LiveBench 64 · ~29 GB | Qwen3.6 27BAlibaba · LiveBench 64 · ~29 GB |
| #4 | Nemotron 3.5 Lightning 30B A3BNVIDIA · LiveBench 44.5 · ~31 GB | Nemotron 3.5 Lightning 30B A3BNVIDIA · LiveBench 44.5 · ~31 GB |
| #5 | Qwen3 14BAlibaba · ~20 GB | Qwen3 14BAlibaba · ~20 GB |
Measured speed
Published output tokens per second, as measured. A speed is compared only when the same checkpoint, quantization, runtime and lengths ran on both machines themselves; same-chip references are not compared.
No published test ran the same model and settings on both machines.
All published measurements · NVIDIA DGX Spark
| gpt-oss 20B MXFP4 MoE | 83.43 ± 0.59 | MXFP4 MoE | CUDA | llama.cpp bench |
|---|---|---|---|---|
| gemma-3-4b-it-qat Q4_0 | 81.05 ± 0.20 | Q4_0 | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Qwen3.5 35B A3B MXFP4 | 63.13 | MXFP4 (MoE; 35B total / 3B active) | llama.cpp CUDA (SM121a, source build) | nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark |
| Qwen3 Coder 30B A3B Q8_0 | 61.06 ± 0.23 | Q8_0 | CUDA | llama.cpp bench |
| Llama 2 7B Q4_0 | 60.39 ± 0.05 | Q4_0 | Vulkan | llama.cpp Vulkan scoreboard · community result |
| gpt-oss-120b MXFP4 MoE | 58.72 ± 0.20 | MXFP4 MoE | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Nemotron-3-Nano-4B Q8_0 | 52.85 ± 0.12 | Q8_0 | llama.cpp CUDA | llama.cpp benches/nemotron · maintainer result |
| GLM-4.7-Flash Q8_0 | 48.68 ± 0.12 | Q8_0 | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Qwen3.8 27B NVFP4 + DSpark | 36.6 | NVFP4 + DSpark | SGLang | NVIDIA Developer Forum |
| Qwen2.5-Coder-7B-Q8_0 | 29.43 ± 0.02 | Q8_0 | llama.cpp CUDA | llama.cpp benches/dgx-spark · maintainer result |
| Nemotron 3 Super 120B A12B Q4_K | 19.94 ± 0.16 | Q4_K (reported as 'Q4_K' by llama-bench; GGUF from ggml-org) | llama.cpp CUDA | llama.cpp benches/nemotron · DGX Spark |
| Qwen3.8 27B NVFP4 + MTP | 18.5 | NVFP4 + MTP (3 speculative tokens) | vLLM official | NVIDIA Developer Forum |
| Qwen3.6 27B Q4_K_M | 11.85 ± 0.01 | Q4_K_M (15.65 GiB) | llama.cpp CUDA (SM121a, source build) | nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark |
| Qwen3.6 27B BF16 | 4.54 ± 0.00 | BF16 (50.10 GiB) | llama.cpp CUDA (SM121a, source build) | nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark |
All published measurements · Framework Desktop 128 GB
| Nemotron 3 Super 120B A12B UD-Q4_K_XL | 14.86 ± 0.06 | UD-Q4_K_XL (llama-bench: Q4_K - Medium, 78.02 GiB) | llama.cpp Vulkan (Mesa RADV, toolbox vulkan_radv) | kyuz0/amd-strix-halo-toolboxes raw llama-bench log (GitHub) · community benchmark |
|---|---|---|---|---|
| Nemotron 3 Super 120B A12B UD-Q4_K_XL | 14.85 ± 0.04 | UD-Q4_K_XL (llama-bench: Q4_K - Medium, 78.02 GiB) | llama.cpp ROCm 7.2.3 (toolbox rocm-7.2.3) | kyuz0/amd-strix-halo-toolboxes raw llama-bench log (GitHub) · community benchmark |
| Qwen3.6 27B UD-Q8_K_XL | 6.51 ± 0.00 | UD-Q8_K_XL (llama-bench: Q8_0, 33.31 GiB) | llama.cpp ROCm 7.2.3 (toolbox rocm-7.2.3) | kyuz0/amd-strix-halo-toolboxes raw llama-bench log (GitHub) · community benchmark |
| Qwen3.6 27B UD-Q8_K_XL | 6.30 ± 0.00 | UD-Q8_K_XL (llama-bench: Q8_0, 33.31 GiB) | llama.cpp Vulkan (Mesa RADV, toolbox vulkan_radv) | kyuz0/amd-strix-halo-toolboxes raw llama-bench log (GitHub) · community benchmark |
Verdict
Both have 128 GB unified memory. By the memory estimate, 849 text models run well on each. NVIDIA DGX Spark lists 273 GB/s of memory bandwidth, 1.1× the 256 GB/s of Framework Desktop 128 GB (a manufacturer spec, not a measured speed). No published test ran the same model and settings on both, so their speeds are not compared here. Launch prices on file: NVIDIA DGX Spark USD 3,999 (2025-10-15) and Framework Desktop 128 GB USD 1,999 (2025-02-25). On these facts neither leads: the same number of text models run well and no shared measurement compares their speed.
Limits and dates
Model counts are memory estimates, not speed measurements or guaranteed fits; verify runtime and checkpoint support. Bandwidth is the manufacturer's figure. Prices are launch or list prices with their dates, not current offers.
Page updated · Model catalog checked · LiveBench 2026-06-25