AI Adoption Studio.
19 Nov 2026
BriefPrivate AI hardware

A smarter model, or more users.

Six setups, from one desk box to an eight-GPU server. Each can run the strongest model that fits, or a fast model for many people. Some figures are measured here. 3 × Spark and several of the stronger models are not.

●measured here ◐sourced ○fits on paper, not tested
01 / 09← → or space · p to print
AI Adoption Studio.
Overview

Strongest model, or most users

DGX Spark

Strongest
Nemotron 120B○
10–30 users
Fastest
Qwen3.6 35B●
30–100 users
Cost 1×

3 × DGX Spark

Strongest
GLM-5.3-Flash across boxes○
a handful
Fastest
3 × Qwen3.6 35B○
100–400 users
Cost 2–4×

GB300

Strongest
GLM-5.3-Flash○
30–100 users
Fastest
Nemotron 120B●
400+ users
Cost 8–25×

RTX ×1

Strongest
Nemotron 120B○
30–100 users
Fastest
Qwen 27B●
30–100 users
Cost 2–4×

RTX ×4

Strongest
GLM-5.3-Flash●
30–100 users
Fastest
4 × Qwen 27B●
100–400 users
Cost 8–25×

RTX ×8

Strongest
DeepSeek V4.1 Flash●
30–100 users
Fastest
8 × Qwen 27B●
400+ users
Cost 8–25×

● measured here · ◐ sourced · ○ fits on paper, not tested. Users = named people doing chat-style work, with the first word in under 2 s for 95% of requests. Bands: a handful (<10), 10–30, 30–100, 100–400, 400+. ○ bands are estimates. Cost = hardware price against one DGX Spark; card prices move fast.

02 / 09Click a setup for details
AI Adoption Studio.
Setup A

DGX Spark

Private assistant for a small office.

Cost 1×
Strongest that fits

Nemotron 3.5 Super 120B○

Users10–30
Throughput34–49 tok/s for one user◐Total not measured. Caution above 32K context.
Good forCareful drafts and summaries for a few people
Fastest useful

Qwen3.6 35B●

Users30–100
Throughput127 tok/s for one user430 tok/s total at 8 users
Good forOffice assistant, document questions, OCR beside it

Nemotron 120B uses 70 of 121 GiB. Qwen3.6 measured on our Spark, Oct 2026. Tier = model size class, not a quality score.

03 / 09← → or space
AI Adoption Studio.
Setup B

3 × DGX Spark

One box per team, or one larger model across three.

Cost 2–4×
Strongest that fits

GLM-5.3-Flash across the boxes○

Or: three Nemotron 120B boxes, one per team○ — 30–100 users, 34–49 tok/s each◐

Usersa handful
ThroughputNot measuredThe network link between boxes slows a split model.
Good forOne larger model shared by a small group
Fastest useful

3 × Qwen3.6 35B○

Users100–400
Throughput127 tok/s for one user~1,290 tok/s total: 3 × one measured box
Good forOne assistant for each team or site

No linked-Spark test exists. GLM-5.3-Flash uses 186 of 363 GiB across three boxes. Tier = model size class, not a quality score.

04 / 09← → or space
AI Adoption Studio.
Setup C

GB300 station

A large model on one GPU, or a fast one for many users.

Cost 8–25×
Strongest that fits

GLM-5.3-Flash 320B○

Users30–100
ThroughputNot measured on this machine186 of 251 GiB; little room left for long contexts.
Good forHarder reasoning and long documents, once tested
Fastest useful

Nemotron 3.5 Super 120B●

Users400+
Throughput622 tok/s for one user12,800 tok/s total at 64 users
Good forMany users at once, bulk generation

Nemotron measured with very short prompts and an engine set to at most 64 requests; real agent prompts will lower it. One GPU: no redundancy. Tier = model size class, not a quality score.

05 / 09← → or space
AI Adoption Studio.
Setup D

RTX PRO 6000 ×1

One card: a 120B model, or a fast team assistant.

Cost 2–4×
Strongest that fits

Nemotron 3.5 Super 120B○

Users30–100
ThroughputNot measured on this cardNVIDIA lists it for one card with 5 draft tokens◐
Good forA stronger assistant for a small team
Fastest useful

Qwen 27B●

Users30–100
Throughput46 tok/s for one user, up to 129 with a draft model1,530 tok/s total at 128 users
Good forTeam assistant, workflow automation

Nemotron 120B uses 70 of 96 GiB. Qwen 27B meets the 2 s target up to 16 users at once; the 128-user total is slower for each user. Tier = model size class, not a quality score.

06 / 09← → or space
AI Adoption Studio.
Setup E

RTX PRO 6000 ×4

A 320B model for a department, or four separate small models.

Cost 8–25×
Strongest that fits

GLM-5.3-Flash 320B●

Users30–100
Throughput190–275 tok/s for one user1,356 tok/s total at 32 users
Good forDepartment assistant with better answers
Fastest useful

4 × Qwen 27B●

Users100–400
Throughput46 tok/s for one user3,800 tok/s total at 128 users
Good forFour customers, one card each

GLM meets the 2 s target up to 8 users at once; Qwen up to 64. Mixed models on the four cards are not measured. Tier = model size class, not a quality score.

07 / 09← → or space
AI Adoption Studio.
Setup F

RTX PRO 6000 ×8

The largest model with 1M context, or hundreds of users on small models.

Cost 8–25×
Strongest that fits

DeepSeek V4.1 Flash●

Users30–100
Throughput107 tok/s for one user630 tok/s total at 8 users
Good forA whole code base or case file in one request
Fastest useful

8 × Qwen 27B●

Users400+
Throughput46 tok/s for one user7,600 tok/s total at 256 users
Good forHosting for many small firms

DeepSeek serves 1M context; prompts up to 144K tested. Qwen meets the 2 s target up to 128 users at once. At high card quotes the cost can pass 25×. Tier = model size class, not a quality score.

08 / 09← → or space
AI Adoption Studio.
Next

Low or irregular use? A cloud API, including EU-hosted ones, is probably cheaper.

Next: test the strongest models on real Czech tasks, then choose the buyer and the use case.

09 / 09Sources: lab reports 16–19 Nov 2026 · Spark evaluations Oct 2026 · NVIDIA model card