For one person chatting with a popular open model, published results so far put the two surprisingly close. On GPT-OSS-120B, an RTX Spark laptop and a DGX Spark both generate somewhere around 55 to 65 tokens per second in the tests we have found. The DGX Spark pulls ahead in what those single-user numbers do not show: documented prompt-processing speed on long inputs, serving several users at once, Linux-only serving engines and clustering for bigger models. The RTX Spark's advantage is that it does all this inside a Windows laptop. We have not run these benchmarks ourselves; every figure below comes from the named source, with its date, and the last section explains how to reproduce a fair test on your own models.
What do published benchmarks show?
| Model (type) | DGX Spark / GB10: generation speed | RTX Spark N1X laptop: generation speed |
|---|---|---|
| GPT-OSS-20B (mixture of experts) | 83.4 tok/s, llama.cpp MXFP4 (llama.cpp DGX Spark benchmark, Feb 2026) | 58 to 59 tok/s, Ollama/llama.cpp, Surface Laptop Ultra 128 GB (StorageReview, Oct 2026) |
| GPT-OSS-120B (mixture of experts) | 58.7 tok/s, llama.cpp MXFP4 (llama.cpp, Feb 2026); 55.4 tok/s (NVIDIA, Oct 2025) | 63.3 tok/s, MSI Prestige N16 Flip AI+ 128 GB; 58.9 tok/s, ASUS ProArt P14 (Techtesters, Oct 2026; tool not stated) |
| Qwen3 30B to 35B (mixture of experts) | 61.1 tok/s, Qwen3-Coder-30B-A3B at Q8 (llama.cpp, Feb 2026) | 78.1 tok/s MSI (Techtesters) and 89.2 tok/s Surface (StorageReview), Qwen3.6-35B-A3B |
| Qwen 27B to 32B (dense) | 9.4 tok/s Qwen3 32B q4 (Ollama, Oct 2025); 12.6 tok/s Qwen3.6-27B NVFP4, vLLM (Openzeka, Jul 2026) | 15.8 tok/s Qwen3.8-27B, MSI (Techtesters) |
| Llama 70B (dense) | 4.4 tok/s q4 (Ollama, Oct 2025); 2.7 tok/s FP8, SGLang (LMSYS, Oct 2025) | No reliable published figure found |
| Prompt processing, GPT-OSS-120B | 2,444 tok/s at 2,048 tokens; 1,567 tok/s at 32K context (llama.cpp, Feb 2026) | Not published for large models |
Single-user (batch 1) results. Tests differ in software version, quantisation, prompt length and date, so treat this as a rough guide rather than a head-to-head. RTX Spark reviews were run on launch drivers, which several reviewers described as unfinished.
Why are the speeds so close?
Because generating text is mostly a memory-bandwidth job. Each new token requires reading the model's active weights from memory, so the speed ceiling is roughly bandwidth divided by the size of those weights. DGX Spark's 128 GB runs at 273 GB/s. NVIDIA has not published an RTX Spark figure, but reviewers at ServeTheHome and StorageReview estimate about 300 GB/s from its LPDDR5X-9400 memory on a 256-bit bus. Two chips with similar bandwidth produce similar single-user speeds.
The table also shows why model type matters more than the machine. Mixture-of-experts models such as GPT-OSS and Qwen3-30B-A3B only use a small part of their weights for each token, so they run fast on both. Dense models such as Llama 70B use every weight every time and run at single-digit speeds on both. If speed matters, choosing the right model is a bigger lever than choosing between these two computers.
On a single chat, the model you pick matters more than which Spark you buy.
Where does DGX Spark pull ahead?
- Long prompts. Reading a long document or a big code base happens in the prefill stage, which depends on compute. DGX Spark's prefill is well documented: llama.cpp's benchmark shows GPT-OSS-120B processing prompts at 2,444 tokens per second, and still 1,567 at a 32K-token context. We found no comparable RTX Spark figure for large models.
- Several users or agents. Batching recovers throughput that one user cannot. In llama.cpp's batched test, GPT-OSS-120B on DGX Spark rose from 57 tokens per second for one request to 353 tokens per second in total across 32. Our post Can One GB10 Support 20 Employees Using AI at the Same Time? covers this.
- Serving engines. vLLM, SGLang and TensorRT-LLM, the engines most used for multi-user serving, are Linux-first. vLLM's documentation says it does not support Windows natively.
- Sustained load. A desk unit on a 240 W supply holds performance for hours. Techtesters found the MSI laptop steady over a 10-minute run, which is encouraging, but StorageReview saw two Surface units cut power under a long all-core CPU test.
- Bigger models. Two DGX Sparks pool 256 GB for models up to about 400B. RTX Spark cannot cluster.
Where does RTX Spark hold its own?
For one user working interactively with models up to the 120B class, the laptop results are competitive, and it delivers them with a screen, keyboard and battery attached. Two practical cautions. The GPU's share of memory is capped by default on the laptops reviewers tested, at about 64 GB on 128 GB machines, so a 60 GB model such as GPT-OSS-120B may need the allocation raised first. And Windows on Arm tool support is still filling in, as our RTX Spark compatibility guide explains.
How can you run a fair test yourself?
Vendor and review numbers are a starting point. Before you buy for a team, test your own model and prompt sizes on both machines with the same method:
- Use the same model file and quantisation on both, for example the same GGUF download.
- Use the same engine and version. llama.cpp builds for both Linux Arm64 and Windows on Arm, which makes it the simplest common ground.
-
Measure both prefill and generation. llama.cpp's
llama-benchreports prompt-processing (pp) and token-generation (tg) speeds; run it with a prompt length that matches your work, such as-p 2048 -n 128. - Test at your real context length, because speed falls as context grows.
- Test concurrency if more than one person will use it, using llama.cpp's batched benchmark or your serving engine's load test.
- Run for at least 10 minutes to see sustained, not burst, performance, and note whether the laptop is plugged in.
- Record software versions and dates. NVIDIA says DGX Spark software updates have improved performance by up to 2.5x since launch, mostly in prefill, so results age quickly.
For a quick check in Ollama, ollama run <model> --verbose prints prompt and generation rates after each reply.
Which should you buy for your models?
- Single user, mixture-of-experts models up to 120B, needs Windows or portability: an RTX Spark laptop such as the MSI Prestige N16 Flip AI+ 128 GB (S$10,854.12) or 64 GB (S$6,331.28).
- Long documents, several users, agents, fine-tuning or models above 120B: a GB10 system such as the NVIDIA DGX Spark (S$8,800) or ASUS Ascent GX10 (from S$8,100).
Prices at SourceIT, listed 28 Oct 2026. For the wider decision, see DGX Spark vs RTX Spark: Same AI Ambitions, Two Very Different Computers and our best laptop to run LLMs locally guide. Apple's alternative is covered in NVIDIA GB10 vs Mac Studio M5 Ultra, and reasoning models in NVIDIA GB10 and DeepSeek.
Buying in Singapore
SourceIT supplies GB10 systems from NVIDIA, ASUS, HP, Dell, Lenovo, GIGABYTE and MSI, and all three MSI Prestige N16 Flip AI+ RTX Spark configurations, with local warranty and a GST invoice. If you want to test a specific model before ordering several units, tell us the model and prompt sizes and we will advise. Browse the NVIDIA DGX Spark collection and MSI NVIDIA RTX Spark AI PCs, or size memory with the AI hardware calculator. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Is RTX Spark faster than DGX Spark for LLMs? For one user on mixture-of-experts models, published results are similar: around 55 to 65 tokens per second on GPT-OSS-120B for both. DGX Spark has better-documented prompt processing and handles several users and Linux serving engines better.
How fast is GPT-OSS-120B on DGX Spark? llama.cpp's DGX Spark benchmark of February 2026 shows 58.7 tokens per second generation and 2,444 tokens per second prompt processing at batch 1, using MXFP4 weights.
How fast is GPT-OSS-120B on an RTX Spark laptop? Techtesters measured 63.3 tokens per second on a 128 GB MSI Prestige N16 Flip AI+ in October 2026, steady over a 10-minute run. The tool and settings were not published.
Why are dense 70B models slow on both? Generation speed is limited by memory bandwidth, and a dense 70B model reads all its weights for every token. On roughly 270 to 300 GB/s of bandwidth that means single-digit tokens per second.
How do I benchmark fairly? Use the same model file, quantisation and engine version on both, measure prompt processing and generation at your real context length, test concurrency if needed, and run for at least 10 minutes.
