RTX Spark vs DGX Spark: What NVIDIA's 120B and 200B AI Model Claims Really Mean

They are memory claims, not speed promises, and both assume heavily compressed models. NVIDIA says DGX Spark can work with "AI models up to 200 billion parameters" and that RTX Spark can "run models up to 120B parameters with 1M context". A 200B-parameter model only fits in DGX Spark's 128 GB if each weight is stored in about 4 bits, and NVIDIA's DGX Spark datasheet ties the figure to FP4 support. Neither claim tells you how fast the model will answer, how long a document it can read at once, or how many people can use it. This post explains each part of the claim, with published figures, so you can judge what a machine will really do with the model you want.

What do NVIDIA and its partners actually claim?

Checked October 2026. NVIDIA's pages do not state the precision behind each parameter count. StorageReview attributes the spread between RTX Spark vendors to different assumptions about bits per weight.

How does a parameter count become gigabytes?

Each parameter is a number the model stores. How much memory it takes depends on precision:

• 16-bit (FP16 or BF16): 2 bytes per parameter, so a 70B model needs about 140 GB.
• 8-bit (FP8): 1 byte, so 70B needs about 70 GB.
• 4-bit (FP4, NVFP4, MXFP4, Q4): about half a byte, so 70B needs about 35 to 43 GB and 200B about 100 GB.

Real files confirm the arithmetic. NVIDIA's 4-bit Llama 3.3 70B is 42.7 GB on Hugging Face and its 8-bit version 72.7 GB; OpenAI's GPT-OSS-120B ships in 4-bit MXFP4 at about 65 GB. So "200B on 128 GB" means 4-bit weights of about 100 GB, leaving roughly 20 to 25 GB for everything else. "120B on 128 GB" is more comfortable, at about 60 to 65 GB of weights.

Lower still is possible. NVIDIA's RTX Spark launch model, Nemotron 3.5 Bolt 74B-A9B, is packed into 24 GB using 2-bit quantisation, according to NVIDIA's developer blog. Each step down in precision saves memory and costs some answer quality, and how much varies by model.

What does "1M context" mean?

Context is how much text the model can consider at once: your question, the documents you paste in and the conversation so far. One million tokens is roughly several thick books. The catch is that context needs memory too. While generating, the model keeps a key-value (KV) cache for every token in the context, and that cache grows with context length and with every simultaneous user. How big it gets depends heavily on the model's design, and NVIDIA does not say which model and settings its 1M claim refers to. Treat it as "possible with a suitable model", not "available with any 120B model".

Long context also slows things down. In llama.cpp's DGX Spark benchmark, GPT-OSS-120B generated 58.7 tokens per second with a short context and 42.8 at a 32K-token context, while prompt processing fell from 2,444 to 1,567 tokens per second.

"Up to 200B" means it fits. It does not mean it flies.

Does a bigger model always run slower?

Not if it is a mixture-of-experts (MoE) model. MoE models hold many parameters but only use a fraction of them for each token. Nemotron Bolt's name, 74B-A9B, says it: 74 billion in total, about 9 billion active. Memory follows the total; speed follows the active part. That is why published results look so uneven:

Single-user results; settings differ between sources. A 120B MoE model can be ten times faster than a 70B dense one on the same box. When a vendor says "runs a 120B model", ask which one. Reasoning models add another twist, because they generate long chains of thought before answering; see NVIDIA GB10 and DeepSeek: How Fast Can It Really Think?

What about the 400B and 700B cluster figures?

These come from pooling memory across GB10 units linked by ConnectX-7: 256 GB for two, 512 GB for four. They are capacity figures too. NVIDIA's own example with two units ran Qwen3-235B at 4-bit with 11.73 tokens per second of generation. Third-party tests show clustering adds capacity far more than speed; we cover the evidence in Two NVIDIA GB10 Systems, One AI Model: Does Performance Actually Double?

How should you read a model-size claim?

1. Ask the precision. 4-bit, 8-bit or 2-bit changes everything.
2. Ask the model. MoE or dense, and which one specifically.
3. Ask the context length the claim was made at, and plan your own.
4. Ask about concurrency. One user, or ten?
5. Check the GPU memory allocation on Windows RTX Spark machines, which reviewers found capped at about 64 GB by default on 128 GB laptops.
6. Leave headroom. Plan on weights plus 20 to 30% for one user, more for a team.

Our AI hardware calculator does the sums for any parameter count, precision and context length, and 128GB RAM Sounds Crazy. Until You Start Running Local AI. explains why memory sets the ceiling.

Buying in Singapore

SourceIT supplies GB10 systems including the NVIDIA DGX Spark (S$8,800) and ASUS Ascent GX10 (from S$8,100), and RTX Spark laptops such as the MSI Prestige N16 Flip AI+ 128 GB (S$10,854.12), with local warranty and a GST invoice. Prices listed 28 Oct 2026. Tell us the model you want to run and we will size the machine. Browse the NVIDIA DGX Spark collection and MSI NVIDIA RTX Spark AI PCs. 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 a DGX Spark really run a 200B model? It can load one at about 4-bit precision, where the weights take roughly 100 GB of its 128 GB. NVIDIA's datasheet ties the figure to FP4 support. Speed depends on the model's design.

What does RTX Spark's "120B with 1M context" mean? It is NVIDIA's claim that a suitable model of up to 120 billion parameters can run with a context window of up to one million tokens. NVIDIA does not state the model or precision behind it.

How much memory does a 70B model need? About 140 GB at 16-bit, 70 GB at 8-bit and 35 to 43 GB at 4-bit, plus room for the KV cache. NVIDIA's 4-bit Llama 3.3 70B file is 42.7 GB.

Why is a 120B model faster than a 70B model on the same machine? If the 120B model is a mixture of experts, it only uses a small share of its weights for each token. GPT-OSS-120B runs at about 59 tokens per second on DGX Spark, while dense Llama 70B runs at 3 to 4.

Does clustering two DGX Sparks double the speed? Not usually. It doubles memory to 256 GB so larger models fit; any speed gain depends on how the model is split. NVIDIA's two-unit Qwen3-235B example generated about 11.7 tokens per second.