Market
Market is the board where Polaris members will sell and give away computers and graphics cards for running models at home. Listings open later. The memory sums and checks below work on any ad you are looking at today.
Listings
Hardware listings open with membership. No listings have been published yet.
Beginner
What a first card has to hold
Video memory (VRAM) is the memory on a graphics card. A language model is a file of weights, the numbers it learned, and those weights sit in memory while it answers. On a desktop PC the fast place is the card's VRAM. Ordinary RAM can hold the overflow, and replies then come more slowly.
Q4_K is a common compressed format. Hugging Face's GGUF notes put it at 4.5 bits per weight. Divide by eight to get bytes. An 8-billion-parameter model needs about 4.5 GB for the weights, and a 3-billion-parameter model about 1.7 GB. The chat needs more on top, and a longer chat needs more, so an 8 GB card can try a small model and a 12 GB card leaves room. At 16-bit, weights take about 2 GB per billion parameters, so that 8-billion model is about 16 GB before the chat starts.
When listings open, filter for at least 12 GB of VRAM, and try a ceiling around 5,000 Swedish kronor for a first card.
A product-sheet photo can be pasted into any ad. Ask for the seller's photo of this card, its power plugs, and a screen that names the chip and the memory. Ollama runs a downloaded model locally and publishes which NVIDIA, AMD, and Apple chips it supports. LM Studio does the same job on the desktop and runs models through llama.cpp. Start around 3 to 8 billion parameters.
Check a first graphics card
I want to run a small AI model on my own computer. I will paste a classified ad for a used graphics card. Using only the numbers in the ad plus any spec page you name, say whether the video memory can hold an 8-billion-parameter model at 4.5 bits per weight (about 4.5 GB for the weights) plus a short chat. Ask me for the card name, the memory in GB, the seller photos, my power supply wattage, and my country. Explain VRAM the first time you use it. Do not invent a price or a test score.
Intermediate
24 GB cards, Macs, and the power plug
Two cards can both say 24 GB and still be a different buy. The GeForce RTX 3090 page lists 24 GB of GDDR6X, and the RTX 4090 page lists 24 GB of GDDR6X too. At 4.5 bits per weight, a 32-billion-parameter model is about 18 GB of weights, so a 24 GB card can hold them plus a chat. The RTX 5090 has 32 GB of GDDR7, which holds the same model with more room for the chat. NVIDIA lists 575 watts total graphics power. A 70-billion-parameter model is about 39 GB of weights, so one card holds only part of the file.
The 4090 spec lists 450 watts total graphics power, an 850 watt system supply, 304 mm length, and three slots. The same page marks NVLink (SLI-ready) as No. The 3090 page describes an NVLink bridge for 30-series cards. Read the partner board's length and plugs against your case and your power supply.
For a Mac, the ad should show the chip and the memory in About This Mac. Apple's Apple silicon session says the CPU and GPU share that memory, which Apple calls unified memory. A Mac at 64 GB or 128 GB can hold the 39 GB file that spills off a 24 GB card.
Filter on memory, power, and shipping. Say how a heavy card is packed. The seller may offer to install Ollama or LM Studio on that computer. Ads for paid services without hardware belong on the jobs board when it opens.
Compare two machines for one model
I will paste two ads: a graphics card and a Mac, or two cards. For each, list memory in GB, the power and size figures I paste, and whether the seller ships. Estimate weight-file size as parameter count times 4.5 divided by 8, and say what extra memory a chat still needs. Compare a 24 GB card with a Mac whose ad shows 64 GB or 128 GB. Quote a source page for any number you add. Point out a missing photo, a missing power figure, or a product-sheet picture standing in for the real item.
Expert
Server hardware compared with the DGX Station
Server buyers read the link between chips and the speed of each memory pool. PCIe is the slot a card uses. DGX Station lists PCIe Gen 5 slots beside a Grace CPU and one Blackwell Ultra GPU. Its NVLink-C2C link runs at 900 GB/s between that CPU and GPU. A card with only PCIe stays on the slower path.
The station's 748 GB is coherent memory at two speeds: 252 GB of HBM3e at 7.1 TB/s on the GPU, and 496 GB of LPDDR5X at 396 GB/s on the CPU. Coherent means both processors share one address space. The fast 252 GB is a bit over ten times a 24 GB card. The other 496 GB is extra room at a lower speed. An ad that says 748 GB of GPU memory has added the pools together.
Match a server ad to a row in Epoch AI's hardware table of published memory and bandwidth, then check the maker's sheet, the wattage, and whether error-correcting (ECC) memory is stated. The station is listed at 1,600 watts and needs a 20 amp circuit. NVIDIA states up to 20 petaFLOPS of AI performance, with FP4 tensor math at 20 petaFLOPS with sparsity and 15 without.
Rented hours, when offered, name the chip, the memory, and the software image, and the machine stays with its owner. A LoRA, extra weights trained on top of a base model, still needs that base model in memory. Weight files stay off this board. Closed ads will show the agreed price after a sale.
Set a server ad next to DGX Station
I will paste a listing for a used accelerator or workstation. Compare its memory, bandwidth, power, interconnect, and any ECC note with NVIDIA DGX Station: 252 GB HBM3e at 7.1 TB/s, 496 GB LPDDR5X at 396 GB/s, NVLink-C2C at 900 GB/s, PCIe Gen 5 slots, up to 20 petaFLOPS, 1,600 W on a 20 A circuit. Keep the fast GPU memory separate from the slower CPU memory. Tell me which new price to look up, and which fields the ad still lacks. Do not invent a street price, a benchmark, or a past sale.
What fits in the memory
These figures are weight-file sizes at Q4_K, using 4.5 bits per weight from the GGUF notes (parameters × 4.5 ÷ 8). They leave out the chat, the app, and the operating system. Measure the real load with the app after you install the card. llama.cpp can run a model that spills out of VRAM by using system RAM as well, and that path is slower.
| Model size | Weights at Q4_K | Where the weights fit |
|---|---|---|
| 3 billion parameters | about 1.7 GB | an 8 GB card, with room |
| 8 billion | 4.5 GB | 8 GB is tight once the chat grows; 12 GB is the safer floor |
| 14 billion | about 7.9 GB | 12 GB holds the weights and a short chat |
| 32 billion | 18 GB | a 24 GB card, or a 32 GB card with room for a longer chat |
| 70 billion | about 39 GB | a Mac with 64 GB or more, or more than one 24 GB card if the software splits the model |
| 16-bit, any size | about 2 GB per billion parameters | an 8-billion model is about 16 GB of weights before the chat |
How a listing will work
When listings open, there will also be a Giveaway tab and an archive of closed sales.
A listing is a machine or a part the buyer can collect or receive by post, shown in the seller's own photos. Write the chip name, the memory in GB, the power figure, what is in the box, the place, and whether you ship. Accounts, API keys, cloud subscriptions, and prompt packs stay off the board. If the shared DGX Station is bought, it is owned by Hrafninn Industries AB and is not sold here. A member's ad is that member's own machine or part.
Filters, once the board is live: memory in GB, category (graphics card, whole PC, Apple silicon Mac, server part), price, place, shipping, and power. Each category needs its own reference photo, separate from any seller's pictures.
Write one price line on every ad you study:
| Line | What you record |
|---|---|
| Asking price | The figure in the ad, the currency, and the date you saw it |
| New price that day | The maker's or a large retailer's price for the same part, dated. A launch price from an old announcement is a different number |
| Memory you actually get | GB, and GB/s if the maker prints it. On the station: see the expert section above |
| Station price | Blank until Polaris publishes the campaign figure. NVIDIA's page says to order through a partner and does not print one public price |
| Price members paid | The agreed price in the archive, after a sale exists. There are no sales yet |
If Polaris buys the shared station, models that do not fit any member's card could run there. About the shared station
Sources
- NVIDIA DGX StationOfficial specifications: 748 GB coherent memory split into 252 GB HBM3e at 7.1 TB/s and 496 GB LPDDR5X at 396 GB/s, NVLink-C2C at 900 GB/s, PCIe Gen 5 slots, up to 20 petaFLOPS, and 1,600 W on a 20 A circuit. The page links the PDF datasheet and tells you to order through a partner.
- GGUF quantization types on Hugging FaceWhat the compressed file formats mean. The table gives Q4_K as 4.5 bits per weight, which is the figure behind the memory sums on this page.
- llama.cppThe local inference project behind many desktop apps. The readme covers integer quantization from 1.5-bit to 8-bit and CPU-plus-GPU runs when a model is larger than video memory. The code is under the MIT license.
- Ollama hardware supportWhich NVIDIA, AMD, and Apple machines current Ollama builds accelerate, including Metal on Apple silicon.
- NVIDIA GeForce RTX 4090 specifications24 GB GDDR6X, 450 W total graphics power, 850 W system supply, 304 mm length, three slots, and NVLink (SLI-ready) marked No.
- NVIDIA GeForce RTX 509032 GB of GDDR7 and 575 W total graphics power.
- NVIDIA GeForce RTX 309024 GB of GDDR6X on the 3090, and the NVLink bridge NVIDIA describes for 30-series cards.
- Explore the new system architecture of Apple silicon MacsApple session transcript: on Apple silicon the CPU and GPU share one memory pool, so a model uses the memory size shown for the Mac.
- Epoch AI, machine learning hardware documentationPublished memory and bandwidth for AI accelerators, with a CSV that Epoch updates. Cite Epoch AI, Data on machine learning hardware, if you reuse the table.