NVIDIA GB10 and DeepSeek: How Fast Can It Really Think?

It depends entirely on which DeepSeek. The small distilled DeepSeek-R1 models run quickly on a GB10 system: published tests put DeepSeek-R1 14B at about 20 tokens per second for one user and 83.5 in total across eight. The 70B distill fits comfortably but runs slowly, at the single-digit speeds typical of dense 70B models. The full 671-billion-parameter DeepSeek-R1 and V3 do not fit in one GB10's 128 GB at all. DeepSeek's newer mixture-of-experts models are where things get interesting: community testers have run DeepSeek-V4-Flash across two and four linked DGX Sparks at usable speeds. We have not benchmarked these models ourselves; every figure below is from the named source. At SourceIT the ASUS Ascent GX10 starts at S$8,100, with the NVIDIA DGX Spark at S$8,800 (listed 28 Oct 2026).

Which DeepSeek models fit on a GB10?

File sizes from Hugging Face model repositories. The distills are smaller models trained to imitate DeepSeek-R1's reasoning and are built on Llama or Qwen base models.

How fast are the distilled models?

DGX Spark results. We found no published DGX Spark benchmark for the 70B distill itself, so the Llama 70B rows indicate the speed class to expect. NVIDIA says DGX Spark software has improved by up to 2.5x since launch, mainly in prompt processing, so these launch-era figures are conservative.

Why does a "thinking" model feel slower?

DeepSeek-R1 and its distills reason step by step before they answer, and that reasoning is generated text. A question that a normal chat model answers in 200 tokens might produce 1,000 to 3,000 tokens of reasoning first. At 20 tokens per second, 2,000 tokens of thinking takes about 100 seconds before the answer appears. At 4 tokens per second, the 70B distill would take over eight minutes. So the useful question is not just "how many tokens per second?" but "how long until I get an answer?" For interactive use, the 14B and 32B distills are the practical choice on a GB10; keep the 70B distill for batch jobs where nobody is waiting.

For a reasoning model, tokens per second is only half the story. The other half is how many tokens it decides to think.

What about the big DeepSeek models?

The full DeepSeek-R1 and V3 at 671B parameters are too large for one or two GB10s, and we found no published result running them on GB10 hardware. DeepSeek's newer mixture-of-experts release, DeepSeek-V4-Flash at 284B parameters, has been tested by the community:

• Two units, llama.cpp: Classmethod (August 2026) found splitting the model across two DGX Sparks gave about 0.97 times the single-unit generation speed and 0.85 times the prompt speed. In other words, clustering added memory, not speed.
• Two units, vLLM: the same tester reached about 76 tokens per second generation with FP8 and FP4 weights.
• Four units with a switch, vLLM with speculative decoding: a report on NVIDIA's developer forum (August 2026) gives 2,582 tokens per second prompt processing and 55.8 tokens per second generation for one request, peaking at about 126 in total with five requests.

These are third-party results with their own settings. They show that frontier-class DeepSeek models are reachable on a small GB10 cluster, at speeds suited to a team of developers rather than a company. How clustering behaves is covered in Two NVIDIA GB10 Systems, One AI Model: Does Performance Actually Double? and Four GB10 Supercomputers on One Desk.

Why run DeepSeek locally at all?

Two reasons come up most often with Singapore organisations. First, data control: running the model on your own hardware means prompts and documents never reach any external service, including DeepSeek's own. Second, cost: a reasoning model generates many tokens per question, which makes per-token cloud pricing expensive for heavy use. Check the licence for your use: DeepSeek published R1 under the MIT licence, and the distilled versions also carry the licence of their Llama or Qwen base model. Some organisations also have policies on which model families staff may use, so confirm internally first.

How to get started

1. Update DGX OS and the inference software first; performance improves with releases.
2. Try DeepSeek-R1 14B or 32B in Ollama or llama.cpp to judge reasoning quality on your own tasks.
3. Measure total time to answer, not just tokens per second.
4. Move to the 70B distill only for tasks where the smaller ones fall short.
5. For many users, serve with a batching engine such as SGLang or vLLM; see Can One GB10 Support 20 Employees Using AI at the Same Time?

Which model sizes suit which hardware is covered in our guide to running LLMs locally.

Buying in Singapore

SourceIT supplies the ASUS Ascent GX10 (1 TB S$8,100, 2 TB S$8,300, 4 TB S$10,635), the NVIDIA DGX Spark (S$8,800) and other GB10 systems, plus the NVIDIA QSFP112 400G DAC cable (S$155) for linking two units, with local warranty and a GST invoice. Browse the NVIDIA DGX Spark collection. 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

Can NVIDIA GB10 run DeepSeek-R1? It runs the distilled DeepSeek-R1 models from 8B to 70B on one unit. The full 671B DeepSeek-R1 does not fit in one unit's 128 GB.

How fast is DeepSeek-R1 14B on DGX Spark? Ollama measured about 20 tokens per second for one user at 4-bit in October 2025, and LMSYS measured 83.5 tokens per second in total across eight requests with SGLang at FP8.

Is DeepSeek-R1 70B usable on a GB10? It fits, at about 42.5 GB at 4-bit, but dense 70B models generate at only about 3 to 4 tokens per second on DGX Spark, which is slow for a reasoning model. It suits batch work better than chat.

Can a GB10 cluster run DeepSeek-V4-Flash? Community testers have run it on two and four linked DGX Sparks, reaching about 76 tokens per second on two units with vLLM and 55.8 for one request on four units.

Is running DeepSeek locally safe for company data? Running it on your own hardware keeps prompts and documents on your network. Check the model licence and your organisation's policy on model use.