# NVIDIA DGX Station GB300 (partner systems) for local AI: current facts

URL: https://insidethemachine.ai/machines/nvidia-dgx-station-gb300 · machines.json 2026-10-26 · prices and speeds checked 2026-10-15 · models.json 2026-10-26

With 252 GB it counts as a 256 GB machine for AI. The smartest model it runs is Qwen3.8-Flash-Next (125B-A6B), 69% as smart as the best paid AI. 101 of 121 chat models fit its 251 GB for AI. Price unknown for this size.

## Specs

- Kind: NVIDIA AI box (Mini PC or AI box). Shape: Desktop. Status: On sale.
- Chip: NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. Memory: 252 GB, Graphics-card memory (VRAM).
- Memory for AI: 251 GB. The share of memory the maker lets the graphics use.
- Memory speed: 7100 GB/s (7.1 TB/s is the GPU's HBM3e. The CPU memory runs at up to 396 GB/s.). Memory type: HBM3E. Spec source: NVIDIA DGX Station product page (Technical Specifications), 2026-10-21, https://www.nvidia.com/en-us/products/workstations/dgx-station/.
- Power: 1600 W (Total system power (20A circuit required); NVIDIA DGX Station product page (Technical Specifications), 2026-10-21, https://www.nvidia.com/en-us/products/workstations/dgx-station/). Launch price: unknown.
- From our data notes: NVIDIA says DGX Station systems ship from ASUS, Dell, GIGABYTE, HP, MSI, Supermicro and Exxact as of 2026-10-21 (NVIDIA blog); a Windows version is expected from partners in December. Ordered through a partner; NVIDIA lists no price. The 748 GB is 252 GB HBM3e GPU memory plus 496 GB LPDDR5X CPU memory (396 GB/s) over NVLink-C2C; fit counts use only the 252 GB. The Gigabyte W775-V10-L01 is one of these systems, listed separately with its store price.

## Memory sizes (Fig. 1)

Rules: We round each machine down to the biggest standard size (8 to 512 GB) whose memory for AI it has. The GeForce RTX 5090 has 31 GB for AI. A standard 32 GB computer has 24 GB and a 48 GB one has 36 GB, so it counts as 32 GB. It counts as that size for the model picks per size and for the recipes. A machine's fit count ("N of 121 chat models fit") uses its own memory for AI, so its machine page, its /machines row, /my and the /best guides give the same number. So 79 of 121 chat models fit the GeForce RTX 5090, and 64 fit a standard 32 GB computer. "The smartest model it runs" uses the machine's own memory for AI, so a 24 GB graphics card (23 GB for AI) can run a bigger model than a 24 GB Mac (18 GB for AI).

| memory GB | for AI GB | counts as GB | chat models that fit its memory for AI (of 121) | for comparison: fit a standard computer of that size (of 121) | smartest model it runs | % of best paid AI | price |
|---|---|---|---|---|---|---|---|
| 252 | 251 | 256 | 101 | 99 | Qwen3.8-Flash-Next (125B-A6B) | 69% | no dated price in our data |

- 252 GB: Why not GLM-5.3-Flash? GLM-5.3-Flash scores higher, but Ollama only offers it in the cloud. On your own computer it runs only with MLX, which needs a Mac.

## The 24/7 helper (Fig. 1a)

Rule: The helper's brain, ears (speech to text) and voice (text to speech) stay loaded together. Ears and voice get up to 15% of the memory for AI each. The brain is the best model that can use tools and fits in the rest. To make an image, the helper pauses the brain and the image model uses its space. Same rule as Fig. 2 on /explore and /my.

| memory GB | brain | ears (speech to text) | voice (text to speech) | free GB | kept for the system GB | image model (the brain pauses) | brain run command |
|---|---|---|---|---|---|---|---|
| 252 | Qwen3.8-Flash-Next (125B-A6B) (140 GB, 69%) | Granite Speech 5.0 2B (6 GB) | Fish Audio Drama 3 (preview) (28 GB) | 77 | 1 | Qwen-Image 2.1 (46 GB) | `ollama run qwen3.8-flash-next:125b-a6b` (Ollama) |


## Top picks per size (Fig. 1b, judged on this machine's own memory for AI)

Fit rule: We count 75% of your memory as usable for AI. The rest is for the system and your apps. Runs well: the model fits in usable memory with at least 10% to spare. Runs, but tight: it fits, but less than 10% of usable memory is left. Close other apps first. Needs N GB+: it does not fit. N is the smallest memory size it fits in. Size unknown: we do not know how much memory it needs. On this page "usable" is the machine's own memory for AI.

- 252 GB (251 GB for AI): smartest it runs Qwen3.8-Flash-Next (125B-A6B), 69%, Runs well, run: `ollama run qwen3.8-flash-next:125b-a6b` (Ollama).
  Top picks: GLM-5.3-Flash 73%, needs 233 GB, needs a Mac; Qwen3.8-Flash-Next (125B-A6B) 69%, needs 140 GB, Runs well; MiMo-V2.6-Flash 66%, needs 216.4 GB, Runs well; Qwen3.8 27B 59%, needs 22.7 GB, Runs well; Nemotron 3.5 Super VL 120B-A12B 46%, needs 105.5 GB, Runs well.
  Best by job, at the standard 256 GB size: code Qwen3.8-Flash-Next (125B-A6B); agents Qwen3.8-Flash-Next (125B-A6B); reads images Qwen3.8-Flash-Next (125B-A6B).
  Set this as your computer: https://insidethemachine.ai/my?os=win&gpu=nvidia-dgx-station-gb300&vram=252

## Prices (Fig. 3, approximate, USD, cheapest first)

Prices are approximate, in US dollars, with the shop and the date we checked. A memory-chip shortage pushed 2026 prices up. Check before you buy.

We found no dated price for the NVIDIA DGX Station GB300 (partner systems) yet.

## Published speeds (Fig. 4, not our tests)

Speeds are copied from published tests by makers and the community, with sources. We did not run them. Different apps and settings give different speeds, so compare with care. On the page each bar is a share of the fastest published result for that model on any machine in our data.

None found yet. Memory speed (7100 GB/s) is the best guide: higher means faster answers.

Check your own computer (read-only, sends nothing): `curl -fsSL https://insidethemachine.ai/scan.sh | sh` (Mac or Linux) or `irm https://insidethemachine.ai/scan.ps1 | iex` (Windows PowerShell).

---

Reference file (https://insidethemachine.ai/machines/nvidia-dgx-station-gb300.md). Its "chat models that fit" counts this machine's own memory for AI.

# NVIDIA DGX Station GB300 (partner systems) for local AI

URL: https://insidethemachine.ai/machines/nvidia-dgx-station-gb300
Type: NVIDIA AI box. Chip: NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. Memory: vram, options 252 GB. Bandwidth: 7100 GB/s (7.1 TB/s is the GPU's HBM3e. The CPU memory runs at up to 396 GB/s.). Spec source: https://www.nvidia.com/en-us/products/workstations/dgx-station/ (2026-10-21).
Price: unknown. Status: current.

## Which models fit

Not a Mac: models marked (Mac only) need MLX on a Mac, so the best it runs is the best with an Ollama download.

| memory GB | usable GB | chat models that fit its own usable GB (of 121) | best 3 that fit | best it runs |
|---|---|---|---|---|
| 252 | 251 | 101 | GLM-5.3-Flash (233 GB, Mac only), Qwen3.8-Flash-Next (125B-A6B) (140 GB), MiMo-V2.6-Flash (216.4 GB) | Qwen3.8-Flash-Next (125B-A6B) |

## Published speeds

None found yet.

Notes: NVIDIA says DGX Station systems ship from ASUS, Dell, GIGABYTE, HP, MSI, Supermicro and Exxact as of 2026-10-21 (NVIDIA blog); a Windows version is expected from partners in December. Ordered through a partner; NVIDIA lists no price. The 748 GB is 252 GB HBM3e GPU memory plus 496 GB LPDDR5X CPU memory (396 GB/s) over NVLink-C2C; fit counts use only the 252 GB. The Gigabyte W775-V10-L01 is one of these systems, listed separately with its store price.

## Questions

**What is the best local LLM for the NVIDIA DGX Station GB300 (partner systems)?**
With 252 GB, Qwen3.8-Flash-Next (125B-A6B) (needs 140 GB). Its Artificial Analysis score is 39.8: 69% of the best paid AI, Claude Opus 5.5. GLM-5.3-Flash scores higher but runs only on a Mac (MLX). 101 of the 121 open chat models fit in its own 251 GB for AI.