Local AI
What will run your model?
Updated
Llama 3.1 8B: 4.5GB to load, 7.5GB for long documents. For long documents, the cheapest card is RTX 5060 Ti 16GB, new $790 Newegg · last read and the cheapest whole machine Mac mini M4, 16GB / 256GB, used $749 used market · last read .
Qwen 3.5 27B: 15.5GB to load, 25.9GB for long documents. For long documents, the cheapest card is Two used RTX 3090s $2,064 eBay · last read and the cheapest whole machine Mac Studio M5 Max, 36GB, new $2,499 Apple · last read . The Nvidia PC pick loads every model in this row, not long documents: RTX 3090 24GB $1,032 eBay · last read .
Llama 3.3 70B: 39.1GB to load, 65.2GB for long documents. For long documents, no card we list is enough; the cheapest whole machine is GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB $3,499 GMKtec · last read . The Nvidia PC pick loads it, not long documents: Two used RTX 3090s $2,064 eBay · last read .
gpt-oss-120b: 65.4GB to load, 109GB for long documents. For long documents, no card we list is enough; the cheapest whole machine is GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB $3,499 GMKtec · last read .
Qwen3-235B-A22B: 131.3GB to load, 218.9GB for long documents. For long documents, no card we list is enough; the cheapest whole machine is Mac Studio M3 Ultra, 512GB, used $17,448 used market · last read .
Above the line, the memory each model needs for long documents (canirun.ai's recommended figure). Choose a size and a solid line marks its weights, what loads it for short prompts, and a dashed line its long-documents figure. Below, the memory a model can use on each route: all of a card's, about two-thirds of a Mac's at 32GB or less and three-quarters above. Labeled prices are the cheapest route at that amount.
The figures as a table
| Model | To load | Long documents | Cheapest card | Cheapest whole machine |
|---|---|---|---|---|
| Llama 3.1 8B | 4.5GB | 7.5GB | RTX 5060 Ti 16GB, new $790 Newegg · last read | Mac mini M4, 16GB / 256GB, used $749 used market · last read |
| Qwen3 14B | 7.8GB | 13GB | RTX 5060 Ti 16GB, new $790 Newegg · last read | Mac Studio M5 Max, 36GB, new $2,499 Apple · last read |
| gpt-oss-20b | 11.7GB | 19.6GB | RTX 3090 24GB, used $1,032 eBay · last read | Mac Studio M5 Max, 36GB, new $2,499 Apple · last read |
| Qwen 3.5 27B | 15.5GB | 25.9GB | Two used RTX 3090s $2,064 eBay · last read | Mac Studio M5 Max, 36GB, new $2,499 Apple · last read |
| Qwen3 30B-A3B | 16.8GB | 27.9GB | Two used RTX 3090s $2,064 eBay · last read | Mac mini M5 Pro, 64GB / 1TB, new $3,199 Apple · last read |
| Llama 3.3 70B | 39.1GB | 65.2GB | none we can source | GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB $3,499 GMKtec · last read |
| gpt-oss-120b | 65.4GB | 109GB | none we can source | GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB $3,499 GMKtec · last read |
| Qwen3-235B-A22B | 131.3GB | 218.9GB | none we can source | Mac Studio M3 Ultra, 512GB, used $17,448 used market · last read |
| DeepSeek R1 671B | 375GB | 624.9GB | none we can source | none we can source |
| Route | Usable | Price |
|---|---|---|
| Mac mini M4, 16GB / 256GB, used | 10.7GB of 16GB | $749 used market · last read |
| Mac mini M6, 16GB / 256GB, new | 10.7GB of 16GB | $899 Apple · last read |
| RTX 5060 Ti 16GB, new | 16GB | $790 Newegg · last read |
| RTX 5070 Ti 16GB, new | 16GB | $1,150 Newegg · last read |
| RTX 3090 24GB, used | 24GB | $1,032 eBay · last read |
| Mac Studio M5 Max, 36GB, new | 27GB of 36GB | $2,499 Apple · last read |
| RTX 5090 32GB, new | 32GB | $6,700 Newegg · last read |
| Two used RTX 3090s | 48GB | $2,064 eBay · last read |
| Mac mini M5 Pro, 64GB / 1TB, new | 48GB of 64GB | $3,199 Apple · last read |
| Mac Studio M5 Max, 40-core GPU, 64GB, new | 48GB of 64GB | about $3,500 Apple · last read |
| Mac Studio M5 Ultra, 96GB, new | 72GB of 96GB | $5,499 Apple · last read |
| Mac Studio M1 Ultra, 128GB, used | 96GB of 128GB | $5,249 used market · last read |
| GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB | 128GB | $3,499 GMKtec · last read |
| Nvidia DGX Spark, 128GB | 128GB | $4,699 reported in our August comparison · last read |
| Framework Desktop, 128GB | 128GB | $3,584 Framework · last read out of stock (Framework's store, rechecked Sept. 16, 2026) |
| Mac Studio M3 Ultra, 256GB, used | 192GB of 256GB | $9,999 used market · last read |
| Mac Studio M3 Ultra, 512GB, used | 384GB of 512GB | $17,448 used market · last read |
Decision guides
- Add a card to a PC you ownBuyIntel Arc Pro B60 24GBfrom $649.99 Newegg · last read
- What each budget buys, from $0BuyRTX 3090 24GBfrom $1,032 eBay · last read
- No desktop, and a 70B or 120B modelBuyGMKtec EVO-X2from $3,499 GMKtec · last read
- Check what your computer runs firstKeepNo price on this answer
- Your exact model, budget and machineToolKeep it, add a card or buy a machineNo price on this answer
Everything in Local AI13
- Keep it, add a card, or buy the whole machine?
- Mac LLM tokens per second: M5 Pro and M6 Mac minis, M5 Max and M1-M4
- AI memory per dollar: Mac Studio, Strix Halo and Nvidia compared
- Can your computer run a local AI model?
- Local AI hardware by budget: four rungs, and the first costs nothing
- Prime Big Deal Days local AI deals: which GPU or mini PC fits your model?
- Prime Big Deal Days 2026: check PC deals before buying
- Which machine to buy for local AI above $4,000
- The cheapest machine for a 70B or gpt-oss-120b model, and which 128GB mini PC to buy
- Which GPU runs your local model: a used RTX 3090, a cheaper 32GB card or nothing
- M5 Pro Mac mini for local AI: 24GB, 48GB or 64GB, and what the same money buys instead
- Buy, lease, rent by the hour or subscribe: the local AI arithmetic
- When a Mac Studio cluster makes sense, and when it doesn't