Can a GB10 Fine-Tune Your Own AI Model, or Is It Mostly for Inference?

It can genuinely fine-tune, within limits. NVIDIA rates DGX Spark for fine-tuning models of up to 70 billion parameters and publishes results: a LoRA fine-tune of Llama 3.1 8B at about 6,970 tokens per second and a QLoRA fine-tune of Llama 3.3 70B at about 760. That is enough to adapt an 8B model to your company's documents in under an hour per pass over a modest dataset, or a 70B model overnight. What a GB10 cannot do is full fine-tuning of large models or training from scratch at any useful scale; that needs data-centre GPUs or a GB300-class system. Most organisations use a GB10 mainly for inference and occasionally for adapting a smaller model. At SourceIT the ASUS Ascent GX10 starts at S$8,100, with the NVIDIA DGX Spark at S$8,800 (listed 28 Oct 2026).

What fine-tuning has NVIDIA published for DGX Spark?

From the table in NVIDIA's developer blog on DGX Spark performance (October 2025), PyTorch, sequence length 2,048. The same article's text quotes different figures; we use the table. A separate PyTorch blog (February 2026) reports a full BF16 fine-tune of Llama 3.1 8B with TorchTune at a 16K context taking about 8 hours per epoch and using about 80% of the 128 GB.

LoRA, QLoRA and full fine-tuning: what is the difference?

• Full fine-tuning updates every weight. It needs memory for the weights, their gradients and optimiser state, typically several times the model's size. On a GB10 that limits it to small models: NVIDIA's example is 3B, and PyTorch's 8B run used most of the memory.
• LoRA freezes the original model and trains small adapter layers on top. It needs far less memory and produces a small file you can swap in and out. It is the usual choice for 8B to 30B models.
• QLoRA does the same with the base model stored in 4-bit, which is how a 70B model fits on a 128 GB system for training.

For most business uses, such as teaching a model your house style, product names, document formats or support answers, LoRA on an 8B to 30B model is the practical starting point.

How long would a real job take?

Using NVIDIA's published throughput, here is the arithmetic for a dataset of 10,000 examples averaging 1,000 tokens each, about 10 million tokens per pass (epoch):

Estimates only: longer sequences, different batch sizes and other models change the figures, and data preparation and evaluation take time too. The point is the scale: an 8B adaptation fits in a lunch break; a 70B one runs overnight.

A GB10 is a fine-tuning workshop, not a training factory.

What tools are supported?

NVIDIA's DGX Spark playbooks include fine-tuning recipes for PyTorch (including across multiple Sparks), NVIDIA NeMo, Unsloth and LLaMA Factory, as well as LoRA fine-tuning for the FLUX image model and for vision-language models. Because DGX OS is the same software family as NVIDIA's data-centre systems, a recipe developed on a GB10 can move to bigger hardware if a job outgrows it.

When is fine-tuning the wrong answer?

More often than vendors suggest. If the goal is for the model to answer from your documents, retrieval-augmented generation (RAG) is usually better: it uses current documents, cites sources, and needs no retraining when a policy changes. Fine-tuning suits teaching a model a style, a format, a vocabulary or a narrow task, not a body of facts that changes. Our post Can NVIDIA GB10 Analyse Thousands of Company Documents Without Sending Them to the Cloud? covers the RAG route.

Can you fine-tune and serve on the same box?

Yes, but not at the same time if the jobs are large. A 70B QLoRA run uses much of the 128 GB, leaving little for a model serving users. Many teams schedule training overnight and serving by day, or keep one GB10 for experiments and another for the shared assistant. Two linked units also allow larger training jobs; NVIDIA's PyTorch playbook covers multi-Spark fine-tuning. See Two NVIDIA GB10 Systems, One AI Model: Does Performance Actually Double?

When do you need more than a GB10?

• Full fine-tuning above about 8B, or full-precision training of 70B models.
• Large datasets, where a 70B QLoRA job would run for days.
• Training from scratch on anything beyond toy scale.

A DGX Station GB300 adds 748 GB of coherent memory and up to 20 PFLOPS of FP4 compute; NVIDIA rates it at 10 PFLOPS FP8, the precision used for much training. See GB10 vs GB300: when to move and our comparison of GB10 vs RTX PRO 6000 Blackwell for workstation training.

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 with local warranty and a GST invoice. For fine-tuning, a 2 TB or 4 TB unit leaves room for datasets and checkpoints. Browse the NVIDIA DGX Spark collection or the AI workstations and servers 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. Universities and research labs may find our research workstation guide useful.

Frequently asked questions

Can DGX Spark fine-tune AI models? Yes. NVIDIA rates it for fine-tuning models up to 70B parameters and publishes results for full fine-tuning of Llama 3.2 3B, LoRA on Llama 3.1 8B and QLoRA on Llama 3.3 70B.

How fast is LoRA fine-tuning on a GB10? NVIDIA's published table shows about 6,970 tokens per second for Llama 3.1 8B LoRA at a 2,048-token sequence length, roughly 24 minutes per pass over 10 million tokens.

Can a GB10 fine-tune a 70B model? With QLoRA, yes. NVIDIA's table shows about 760 tokens per second for Llama 3.3 70B QLoRA, which suits overnight jobs on modest datasets.

Should I fine-tune or use RAG for company documents? Usually RAG, which answers from current documents and cites them. Fine-tune to teach a style, format or narrow task.

Which fine-tuning tools run on DGX Spark? NVIDIA's playbooks cover PyTorch, NVIDIA NeMo, Unsloth and LLaMA Factory, plus LoRA for FLUX image models and vision-language models.