NVIDIA · GB10 · AI box

Best AI Models for NVIDIA DGX Spark (128 GB)

Runs 14B and 120B MoE models well; 27–31B and 70B fit but are slow (10 and 4 tok/s).

128 GB unified120 GB for models273 GB/s125 TFLOPS FP16240 W2025

GB10 from

ASUS Ascent GX10$2,999 DGX Spark$3,999

What LLMs can the NVIDIA DGX Spark run?

#	Model	Speed	Quant · Memory
1	Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 54.0 t/s	54.0 tok/s	Q4_K_M · 22.0 GB
2	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 58.4 t/s	58.4 tok/s	Q4_K_M · 16.5 GB
3	Qwen3 30B A3B Thinking 2507Alibaba Q4_K_M · 19.4 GB · 70.3 t/s	70.3 tok/s	Q4_K_M · 19.4 GB
4	GLM 4.5 AirZ.ai Q4_K_M · 67.6 GB · 18.7 t/s	18.7 tok/s	Q4_K_M · 67.6 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 90.9 t/s	90.9 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · Memory
1	Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 54.0 t/s	54.0 tok/s	Q4_K_M · 22.0 GB
2	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 58.4 t/s	58.4 tok/s	Q4_K_M · 16.5 GB
3	GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 62.8 t/s	62.8 tok/s	Q4_K_M · 13.1 GB
4	GPT OSS 120BOpenAI Q4_K_M · 70.8 GB · 46.4 t/s	46.4 tok/s	Q4_K_M · 70.8 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 90.9 t/s	90.9 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · Memory
1	Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 54.0 t/s	54.0 tok/s	Q4_K_M · 22.0 GB
2	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 58.4 t/s	58.4 tok/s	Q4_K_M · 16.5 GB
3	Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 27.9 t/s	27.9 tok/s	Q4_K_M · 6.6 GB
4	GPT OSS 120BOpenAI Q4_K_M · 70.8 GB · 46.4 t/s	46.4 tok/s	Q4_K_M · 70.8 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 90.9 t/s	90.9 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · Memory
1	Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 54.0 t/s	54.0 tok/s	Q4_K_M · 22.0 GB
2	Qwen3.8 27BAlibaba Q4_K_M · 17.9 GB · 10.2 t/s	10.2 tok/s	Q4_K_M · 17.9 GB
3	Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 10.2 t/s	10.2 tok/s	Q4_K_M · 17.9 GB
4	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 90.9 t/s	90.9 tok/s	Q4_K_M · 19.3 GB
5	GLM 4.7 FlashZ.ai Q4_K_M · 19.9 GB · 52.4 t/s	52.4 tok/s	Q4_K_M · 19.9 GB

#	Model	Speed ↓	Quant · Memory
1	Qwen3 30B A3B Thinking 2507AlibabaHow to run → Q4_K_M · 19.4 GB · 70.3 t/s	70.3 tok/s	Q4_K_M · 19.4 GB
2	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 90.9 t/s	90.9 tok/s	Q4_K_M · 19.3 GB
3	Qwen3 30B A3B Instruct 2507Alibaba Q4_K_M · 19.4 GB · 70.3 t/s	70.3 tok/s	Q4_K_M · 19.4 GB
4	GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 62.8 t/s	62.8 tok/s	Q4_K_M · 13.1 GB
5	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 58.4 t/s	58.4 tok/s	Q4_K_M · 16.5 GB

#	Model	Speed	Agents Score	Quant · Memory
1	GPT OSS 120BOpenAIHow to run → Q4_K_M · 70.8 GB · 46.4 t/s	46.4 tok/s		Q4_K_M · 70.8 GB
2	Qwen3.8 27BAlibaba Q4_K_M · 17.9 GB · 10.2 t/s	10.2 tok/s		Q4_K_M · 17.9 GB
3	Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB · 54.0 t/s	54.0 tok/s	overall	Q4_K_M · 22.0 GB
4	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 90.9 t/s	90.9 tok/s		Q4_K_M · 19.3 GB
5	GLM 4.7 FlashZ.ai Q4_K_M · 19.9 GB · 52.4 t/s	52.4 tok/s		Q4_K_M · 19.9 GB

See all 401 models on the NVIDIA DGX Spark →Bars = traction for models without a Score yet

About the NVIDIA DGX Spark

The NVIDIA DGX Spark is a 128 GB AI box with more memory than speed: it runs 14B and 120B MoE models well. DGX OS keeps about 8 GB, leaving 120 GB for models. With 125 TFLOPS it reads long prompts faster than most 128 GB machines: a 4K-token prompt takes 690 ms. At its $3,999 MSRP it gives 13.5 tok/s per $1,000, 6th of the 8 machines compared here. GB10 (Blackwell), released 2025.

Speed in 128 GB#6273 GB/s

Prompt reading#2125 TFLOPS

For models120 GBDGX OS keeps 8 GB

ChipGB10released 2025

NVIDIA DGX Spark strengths and weaknesses

Strengths

Runs 120B MoE models: GPT OSS 120B at 46.4 tok/s in 70.8 GB
Reads long prompts fast: A 32K-token prompt on Qwen3.6 35B A3B starts answering in 5.5 s
FP8 and NVFP4 run natively: Blackwell takes FP8 and NVFP4 checkpoints in vLLM as published

Weaknesses

Slow on large dense models: Llama 3.3 70B Instruct at 3.8 tok/s, Qwen3.8 27B at 10.2 tok/s
Same chip for less: ASUS Ascent GX10 uses the same GB10 at $2,999 MSRP

Memory for models

120 GB of 128 GB

35B MoE22 GB70B51.3 GB120B MoE70.8 GB235B119.4 GB

35B MoE22 GB120B MoE70.8 GB235B119.4 GB

For models 120 GBDGX OS keeps 8 GBModel size at Q4

Everything on the size ladder fits. Qwen3.6 35B A3B and Gemma 4 26B A4B IT together use 38.5 GB and leave 81.5 GB free.

Speed per dollar and per watt

NewUsed

On Qwen3.6 35B A3B (35B params, 3B active) at Q4_K_M. New means a shop price, either a street quote or the maker's own price; used is compared with used. The last two columns are the speed you get per $1,000 of price and per 100 W of power; higher is better. The NVIDIA DGX Spark has no current new price, so it is shown at its MSRP.

Machine	New price	tok/s	tok/s per $1,000	tok/s per 100 W
RTX 509032 GB	about $5,376	213	39.7	37.1
Framework Desktop Ryzen AI Max+ 395128 GB	about $1,999	46.7	23.4	33.4
Mac Studio M4 Max128 GB	about $3,499	64.2	18.3	44.3
ASUS Ascent GX10128 GB	about $2,999	54.0	18.0	–
Mac Studio M2 Ultra192 GB	about $5,599	75.9	13.6	25.7
DGX Spark128 GB	about $3,999	54.0	13.5	22.5
MacBook Pro 16" M5 Max128 GB	about $5,399	67.8	12.6	–
Mac Studio M3 Ultra256 GB	about $7,099	76.6	10.8	28.4

Machine	Used price	tok/s	tok/s per $1,000	tok/s per 100 W
RTX 409024 GB	about $3,000	151	50.3	33.6
RTX 509032 GB	about $6,400	213	33.3	37.1

Cost per token: the NVIDIA DGX Spark against 24–256 GB machines and GPUs

On Qwen3.6 35B A3B (35B params, 3B active, Q4_K_M, 4K context). Speed, cost per token and power per token depend on the model.

24–36 GB96–128 GB192+ GBthis AI boxat MSRP

Cost per 1M output tokenson Qwen3.6 35B A3B · card over 3 years + power

$4.88

Machine, MSRP: $4.70
Power: $0.19

3-year cost of ownership$4,157

Or rent an A100 in the cloud$1.11/hr

Renting costs as much as buying after ≈3,600 hours (≈2.5 years at 4 h a day).

Include card price

Power$/kWhUseh/dayKeepyears

Each dot is a machine or GPU; hover or tap it for its numbers, click to open it. Machines without a price for the chosen basis are left out. Machines without a published power figure are left out of Per watt and costed without power. How we calculate cost per token

Similar devices and alternatives

The same chip from other makers first, then machines and GPUs people weigh against it, each with the one reason it is worth a look. Speeds on Qwen3.6 35B A3B.

Same chip and memory for $1,000 lessASUS Ascent GX10ASUS · 119 GB for models · GB10

same speed▼ $1,000 less▼ 1 GB$2,999

Same family 19% faster, 4.5 GB more memoryMac Studio M4 Max (128 GB)Apple · 124.5 GB for models · M4 Max

▲ 19% speed▲ 4.5 GB$3,499

26% faster, 4.5 GB more memoryMacBook Pro 16" M5 Max (128 GB)Apple · 124.5 GB for models · M5 Max

▲ 26% speed▲ 4.5 GB$5,399

About as fastFramework Desktop Ryzen AI Max+ 395Framework · 96 GB for models · Ryzen AI Max+ 395

▼ 13% speed▼ 24 GB$1,999

2.8× faster if your models fit in 24 GBRTX 4090NVIDIA · 24 GB VRAM · GPU

▲ 180% speed▼ 96 GB$1,599rent $0.34/hr

4.0× faster if your models fit in 32 GBRTX 5090NVIDIA · 32 GB VRAM · GPU

▲ 295% speed▼ 88 GB$5,376rent $0.50/hr

NVIDIA DGX Spark tokens per second by model size

Output speed depends on memory bandwidth (273 GB/s). Grey rows need more than the 120 GB available.

9BQwen3.5 9B · Q4_K_M · 6.6 GB
27.9tok/s
14BQwen3 14B · Q4_K_M · 11.9 GB
18.7tok/s
27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
10.2tok/s
35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
54.0tok/s
70BLlama 3.3 70B Instruct · Q4_K_M · 51.3 GB slow
3.8tok/s
120B MoEGPT OSS 120B · Q4_K_M · 70.8 GB
46.4tok/s

Estimated speeds: Q4_K_M, 4K context, one user. How we estimate

Unified-memory platforms: writing vs compute

Writing, tok/sCompute, TFLOPS

Orin64 GB · 204.8 GB/s
41.7
43
Ryzen AI Max+ 395128 GB · 256 GB/s
46.7
30
GB10128 GB · 273 GB/s
54.0
125
Jetson T5000 (Thor)128 GB · 273 GB/s
54.0
259
M4 Pro64 GB · 273 GB/s
43.2
17

Machines with about the same memory bandwidth write at about the same speed; compute decides how fast prompts are read. Speeds on Qwen3.6 35B A3B.

All 1290 models on the NVIDIA DGX Spark

Featured only

Best quant that fits in 120 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works

LLM models compatible with NVIDIA DGX Spark — ranked by Score, recency and Fit
Model	Quant	VRAM	Speed	Context	Fit
Qliphoth 12B V1.212.2BChat Q4_K_M·22.0 t/s	Q4_K_M	11.0 GB9%	22.0 t/s	1024K	B62, Decent
Scout 4B4.3BChatReasoningRoleplay Q4_K_M·54.8 t/s	Q4_K_M	3.6 GB3%	54.8 t/s	131K	A82, Runs well
H2o Danube3 500M Chat514MChat Q4_K_M·249.9 t/s	Q4_K_M	1.0 GB1%	249.9 t/s	8K	S100, Runs great
Gemma3 1B IT1BChat BF16·80.7 t/s	BF16	3.6 GB3%	80.7 t/s	—	S88, Runs great
ERNIE 4.5 0.3B Paddle361MChat BF16·170.6 t/s	BF16	1.2 GB1%	170.6 t/s	131K	S98, Runs great
MedScholar 1.5B1.5BChat Q4_K_M·138.6 t/s	Q4_K_M	1.7 GB1%	138.6 t/s	33K	S96, Runs great
CHADROCK3.6 35B UNCENSORED MTP STRIX LEAN35BChat BF16·2.3 t/s	BF16	80.9 GB67%	2.3 t/s	—	C38, Usable
EXAONE 3.5 7.8B Instruct7.8BChat Q4_K_M·34.4 t/s	Q4_K_M	7.7 GB6%	34.4 t/s	33K	A73, Runs well
Meme Trix MoE 14B A8B V113.7BChat Q4_K_M·32.6 t/s	Q4_K_M	10.7 GB9%	32.6 t/s	1073K	A72, Runs well
DeepSeek V4 Flash JANG CRACK33.5BChat Q4_K_M·8.7 t/s	Q4_K_M	20.6 GB17%	8.7 t/s	1049K	B46, Decent
Gpt2 Mini39MChat Q4_K_M·5915.0 t/s	Q4_K_M	0.0 GB0%	5915.0 t/s	1K	S100, Runs great
NVIDIA Nemotron 3 Nano 4B BF164.0BChatFunctions Q4_K_M·61.2 t/s	Q4_K_M	2.9 GB2%	61.2 t/s	262K	A84, Runs well
Ornith 1.0 9B9.4BChatFunctions Q4_K_M·28.6 t/s	Q4_K_M	6.5 GB5%	28.6 t/s	262K	A69, Runs well
Qwen2.5 1.5B Instruct1.5BChatFunctions Q4_K_M·138.6 t/s	Q4_K_M	1.7 GB1%	138.6 t/s	33K	S96, Runs great
Qwen3 8B 64k Context 2X Josiefied Uncensored8.2BChatReasoningFunctions Q4_K_M·32.1 t/s	Q4_K_M	7.6 GB6%	32.1 t/s	66K	A71, Runs well

NVIDIA DGX Spark specs and power

Specs

Chip: GB10 · Blackwell GPU (GB10)
CPU cores: 20
GPU cores: 6,144
Unified memory: 128 GB LPDDR5X
For models: 120 GB (DGX OS keeps 8 GB)
Memory bandwidth: 273 GB/s
FP16 (dense): 125 TFLOPS
Power: 240 W max
Size: 150x150x50.5mm, 1.2kg
Networking: ConnectX-7 200GbE
OS: DGX OS (Ubuntu-based)
Released: 2025-10-15
MSRP: $3,999

NVIDIA DGX Spark FAQ

Is the NVIDIA DGX Spark worth it for local LLMs in 2026?

At its $3,999 MSRP it gives 13.5 tok/s per $1,000 on Qwen3.6 35B A3B, 6th of the 8 machines compared here. It runs 14B and 120B MoE models well; 27–31B and 70B fit but are slow (10 and 4 tok/s).

How much memory is available for AI on the NVIDIA DGX Spark?

About 120 GB of the 128 GB. DGX OS keeps roughly 8 GB for itself and other apps.

What is the largest model the NVIDIA DGX Spark can run?

Qwen3 235B A22B Thinking 2507 (235.1B) is the largest at 4-bit or better: at Q4_0 it uses 119.4 GB, 100% of the available memory, and runs at about 14.7 tok/s.

Can the NVIDIA DGX Spark run Qwen3.6 35B A3B?

Yes. Qwen3.6 35B A3B fits at Q4_K_M, using 22.0 GB of the 120 GB available, and it is the best model that runs well on this machine.

Can the NVIDIA DGX Spark run 70B models?

Yes, but slowly: Llama 3.3 70B Instruct at Q4_K_M uses 51.3 GB and writes about 3.8 tok/s.

How long does the NVIDIA DGX Spark take to read a long prompt?

On Qwen3.6 35B A3B, about 690 ms for a 4K-token prompt and 5.5 s for a 32K-token document before the first word appears.

Is renting a GPU cheaper than buying the NVIDIA DGX Spark?

An A100 (80 GB) rents from $1.11/hr. Renting costs as much as buying after ≈3,600 hours (≈2.5 years at 4 h a day).

48 GB GPUs AI boxes Mac Pro M2 Ultra Mac Studio M1 Ultra Mac Studio M2 Ultra Mac Studio M5 Max MacBook Pro 14" M5 Max Jetson AGX Thor Developer Kit Best coding LLMs All hardware