The QNAP QAI-h1290FX takes one NVIDIA card, and QNAP validates two: the RTX PRO 4500 with 32 GB and the RTX PRO 6000 Max-Q with 96 GB. The choice decides which models you can run and how many people can talk to them at once.
The other parts of the box are the same either way: AMD EPYC 7302P, 128 GB ECC memory, twelve SSD bays and a 750 W supply. Only the card changes. The right one depends on two questions: how many people, and how large a model.
The two cards side by side
- RTX PRO 4500 Blackwell: 32 GB, 200 W, memory bandwidth 896 GB/s. This is our Office AI package.
- RTX PRO 6000 Blackwell Max-Q: 96 GB, 300 W, memory bandwidth 1,792 GB/s. This is our Company AI package.
Memory size decides what fits. Memory bandwidth, how fast the card reads its own memory, decides how fast words appear, because for every word the card reads the active part of the model once. A dense model uses all its weights for every word, so it follows bandwidth closely. The 96 GB card has twice the bandwidth and draws 100 W more.
What the numbers say, model by model
Q4 is a 4-bit model file; Q8 is 8-bit, with higher quality and about twice the size. The selector picks the best one that fits each card. Speed is tok/s (tokens per second, roughly words per second) for one user at a 32,000-token context.
| Model | Type | 32 GB card | 96 GB card |
|---|---|---|---|
| Nemotron SEA-LION 30B-A3B | MoE | Q4, about 145 tok/s | Q8, about 176 tok/s |
| Gemma 4 26B-A4B | MoE | Q4, about 132 tok/s | Q8, about 159 tok/s |
| gpt-oss-20b | MoE | Q8, about 145 tok/s | Q8, about 200 tok/s |
| Qwen3.8 27B | Dense | Q4, about 35 tok/s | Q8, about 42 tok/s |
| Gemma 4 31B | Dense | Q4, about 30 tok/s | Q8, about 37 tok/s |
| gpt-oss-120b | MoE | Partly in system RAM, about 31 tok/s | Q8, about 175 tok/s |
| Qwen3.5 122B-A10B | MoE | Partly in system RAM, about 14 tok/s | Q4, about 129 tok/s |
| Llama 3.3 70B | Dense | Partly in system RAM, about 3 tok/s | Q4, about 26 tok/s |
MoE means mixture-of-experts: the model has many weights but uses only a few billion for each word, so it runs fast. A dense model uses everything, so it is slower. Two things stand out. First, the dense models are not twice as fast on the bigger card, even though its bandwidth is double, because they run at Q8 there and the file is about twice as large. Compare a like-for-like case, gpt-oss-20b at Q8 on both cards: about 145 against about 200 tok/s. Second, the MoE models are quick on both cards. Everything that fits fully in VRAM runs faster than most people read; only the offloaded rows on the 32 GB card slow down.
How many people can chat at once
Each conversation needs its own KV cache, the model's working memory. At a 32K context that cache is large, so memory, not speed, limits the number of people. The chart shows how many 32K conversations fit in memory for five models.
Conversations at 32K context that fit in memory
Estimate. Higher is more people at once. Replies slow down when many people type in the same moment.
Source: EMD Lab AI selector estimate, calculated from model files and configuration files. Not a measurement. The 96 GB card is shown with 8-bit models, the 32 GB card with 4-bit models (gpt-oss-20b is 8-bit on both).
The 96 GB card also holds more at once: the same model at higher quality, and still room for many chats. Not everyone types at the same moment, so these counts support a larger group than the number suggests. Treat them as an upper limit for memory, not a promise.
Choose the 32 GB card (Office AI) if
Your team is small, up to about 30 people, and your main jobs are document questions, summaries, drafting and meeting minutes. A 27B to 31B dense model at Q4, or a fast MoE model such as Nemotron SEA-LION 30B-A3B or Gemma 4 26B-A4B, covers that well. Indonesian-language work is a good fit for the SEA-LION model.
What it will not do: hold the 70B and 120B class (gpt-oss-120b, Llama 3.3 70B, Qwen3.5 122B-A10B) fully in VRAM. They run with part of the model offloaded to the box's system RAM: gpt-oss-120b is still usable at an estimated 31 tok/s, but the dense Llama 3.3 70B drops to about 3 tok/s. Many users on a dense 27B model will also hit the memory limit quickly, about five long conversations.
Choose the 96 GB card (Company AI) if
Your team is larger, about 30 to 100 people with only some of them active together, or you want 8-bit quality, or a 120B-class model such as gpt-oss-120b for stronger reasoning. It also gives you room to run the assistant, Whisper transcription and an image model side by side.
What it will not do: match the best cloud models on hard reasoning, take a second card (QNAP lists no two-card setup), or use the 600 W Workstation Edition, which does not fit this box. Dense 70B models such as Llama 3.3 70B stay slow, about 26 tok/s.
The short version
- 32 GB: small team, 27B to 31B dense models at Q4 and fast MoE models. Lower power at 200 W.
- 96 GB: bigger team, 8-bit quality, 120B-class models and many more conversations at once. 300 W.
- Bandwidth (896 against 1,792 GB/s) explains speed on dense models; memory explains the number of users.
- If you are close to the line, pick by headcount and by whether you want a 120B-class model. Both cards use the same slot, so moving up later is a card swap, not a new box.
All speeds and conversation counts are estimates calculated by the selector. The QNAP AI server page lists all models for both cards, and you can try your own in the model finder.