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.
Strongest model, or most users
DGX Spark
3 × DGX Spark
GB300
RTX ×1
RTX ×4
RTX ×8
● 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.
DGX Spark
Private assistant for a small office.
Cost 1×Nemotron 3.5 Super 120B○
Qwen3.6 35B●
Nemotron 120B uses 70 of 121 GiB. Qwen3.6 measured on our Spark, Oct 2026. Tier = model size class, not a quality score.
3 × DGX Spark
One box per team, or one larger model across three.
Cost 2–4×GLM-5.3-Flash across the boxes○
Or: three Nemotron 120B boxes, one per team○ — 30–100 users, 34–49 tok/s each◐
3 × Qwen3.6 35B○
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.
GB300 station
A large model on one GPU, or a fast one for many users.
Cost 8–25×GLM-5.3-Flash 320B○
Nemotron 3.5 Super 120B●
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.
RTX PRO 6000 ×1
One card: a 120B model, or a fast team assistant.
Cost 2–4×Nemotron 3.5 Super 120B○
Qwen 27B●
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.
RTX PRO 6000 ×4
A 320B model for a department, or four separate small models.
Cost 8–25×GLM-5.3-Flash 320B●
4 × Qwen 27B●
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.
RTX PRO 6000 ×8
The largest model with 1M context, or hundreds of users on small models.
Cost 8–25×DeepSeek V4.1 Flash●
8 × Qwen 27B●
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.
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.