DGX SparkOut of stock
Lenovo ThinkStation PGX
1TB
$5,859at B&H Photo
128 GB · 273 GB/s
Fits~5 tok/sQ4_K_M
Check priceWill it run? · Llama 3.3 70B on DGX Spark
Can DGX Spark run Llama 70B?
Yes, it fits easily, but it's slow. Llama 3.3 70B needs about 46 GB at Q4_K_M, and a DGX Spark has ~109 GB usable of its 128 GB. Because it's a dense model, every token reads all 70B weights through 273 GB/s of memory, so we estimate only ~5 tok/s at 4-bit (3 tok/s at Q8_0, which also fits at ~80 GB). The one public 70B run we found on a GB10 box measured 4.7 tok/s (Dell Pro Max GB10, Ollama (Geerling)), against our 4.9. Prices & stock checked 2026-10-24. Prices change daily and some retailers block our checker, so always confirm on the store page.
The numbers
| Q4_K_M (4-bit) | Q8_0 (8-bit) | |
|---|---|---|
| Memory needed | ~46 GB | ~80 GB |
| DGX Spark usable | ~109 GB | ~109 GB |
| Fit | Fits~5 tok/s | Fits~3 tok/s |
Memory = weights + 1.5 GB + 5% of weights (~8K context, one user). Speed = 1 ÷ (active GB ÷ (77% × 273 GB/s) + per-token overhead). How we estimate
Measured 70B runs vs our formula
| Hardware | Model / run | Measured tok/s | Our formula | Source |
|---|---|---|---|---|
| Strix Halo256 GB/s | Llama 3.3 70B Q4_K_Mllama.cpp Vulkan tg128 | 4.8 | 4.8 | source |
| M3 Ultra 80c819 GB/s | Llama 3.3 70B 4-bitMLX (oMLX) 1K ctx | 17.0 | 16.8 | source |
| RTX PRO 60001,792 GB/s | Llama 3.3 70B Q4_K_Mllama.cpp tg128 | 34.0 | 33.8 | source |
| Framework Desktop (Strix Halo)256 GB/s | Llama 3.1 70B Q4_K_MOllama (Geerling) | 5.0 | 4.8 | source |
| Dell Pro Max GB10273 GB/s | Llama 3.1 70B Q4_K_MOllama (Geerling) | 4.7 | 4.9 | source |
| M3 Ultra 512GB819 GB/s | Llama 3.1 70B Q4_K_MOllama (Geerling) | 14.1 | 15.7 | source |
Public single-user decode results we use to calibrate the formula. Full list: our estimates vs 52 real benchmarks.
Llama 70B on DGX Spark vs other rigs
| Machine | Price | Bandwidth | 4-bit fit & est. speed |
|---|---|---|---|
| PNY NVIDIA RTX PRO 6000 BlackwellRTX PRO 6000 (card only, needs a host PC) | $17,899 | 1,792 GB/s | Fits~34 tok/s |
| Mac Studio M5 UltraMac Studio | $5,499 | 1,200 GB/s | Fits~23 tok/s |
| Lenovo ThinkStation PGXDGX Spark / GB10 | $5,859 | 273 GB/s | Fits~5 tok/s |
| GMKtec EVO-X2Strix Halo | $2,200 | 256 GB/s | Tight fit~5 tok/s |
Cheapest priced config per platform that fits it, fastest first. Dense 70B speed tracks memory bandwidth almost exactly. More: best hardware for Llama 3.3 70B.
What the Spark is better at
The Spark's 128 GB is better spent on big mixture-of-experts models, which only read a fraction of their weights per token:
- Qwen3.5 122B-A10B: 122B total but only 10B active, ~79 GB, ~31 tok/s (Q4_K_M)
- Nemotron 3 Super 120B-A12B: 120B total but only 12B active, ~78 GB, ~26 tok/s (Q4_K_M)
- GPT-OSS 120B: 117B total but only 5.1B active, ~67 GB, ~61 tok/s (MXFP4 (native))
- Llama 4 Scout 109B-A17B: 109B total but only 17B active, ~71 GB, ~19 tok/s (Q4_K_M)
DGX Spark and GB10 boxes we track
DGX SparkOut of stock
ASUS Ascent GX10
1TB
$5,999at B&H Photo
128 GB · 273 GB/s
Fits~5 tok/sQ4_K_M
Check price
DGX SparkAvailable
NVIDIA DGX Spark
Founders Edition, 4TB
$6,950at NVIDIA Marketplace
128 GB · 273 GB/s
Fits~5 tok/sQ4_K_M
Check price
DGX SparkOut of stock
HP ZGX Nano G1n
4TB
$7,699at B&H Photo
128 GB · 273 GB/s
Fits~5 tok/sQ4_K_M
Check price
DGX SparkStock unknown
Gigabyte AI TOP ATOM
4TB
$8,499at B&H Photo
128 GB · 273 GB/s
Fits~5 tok/sQ4_K_M
Check priceFAQ
Can DGX Spark run Llama 3.3 70B?
Yes. At Q4_K_M it needs about 46 GB and the Spark has ~109 GB usable, so it fits with room for longer context. Expect roughly 5 tokens/sec (our estimate) for one user.
How fast is Llama 70B on DGX Spark?
About 5 tok/s at 4-bit and 3 tok/s at 8-bit by our bandwidth formula. A public Dell Pro Max GB10 run measured 4.7 tok/s. That's readable but slow for chat; prompt processing is where the Spark is strong.
What's the cheapest DGX Spark-class box?
Of the GB10 boxes we track, the Lenovo ThinkStation PGX at $5,859 (B&H Photo, checked 2026-10-24). They all share the same GB10 chip, 128 GB and 273 GB/s, so 70B speed is the same on each.
Can DGX Spark run Llama 70B at 8-bit?
Yes, Q8_0 needs about 80 GB of the ~109 GB usable, but speed drops to ~3 tok/s.
What runs better than Llama 70B on DGX Spark?
Mixture-of-experts models only read their active weights per token, so they're far faster: Qwen3.5 122B-A10B ~31 tok/s; Nemotron 3 Super 120B-A12B ~26 tok/s; GPT-OSS 120B ~61 tok/s; Llama 4 Scout 109B-A17B ~19 tok/s.