Can I run it?
Choose your hardware and see which open models and quantizations fit, at which context length.
What you enter stays on your computer.
On GeForce RTX 5090, 32 GB, 736 models fit at 8,192 tokens or more.
| Model | Quantization | Weights | Longest context that fits | Memory then | Where | Speed, estimate |
|---|---|---|---|---|---|---|
| Llama-4-Scout-17B-16E-Instruct | IQ2_XXS | 28.1 GiB | 8,192 | 31.5 GiB | On the device | 135 to 217 tokens a second |
| Ling-flash-2.0 | IQ2_XXS | 25.3 GiB | 32,768 | 29.0 GiB | On the device | 238 to 380 tokens a second |
| Solar-Open-100B | IQ2_XXS | 25.3 GiB | 16,384 | 30.1 GiB | On the device | 140 to 224 tokens a second |
| Qwen3-Next-80B-A3B-Instruct | Q2_K | 27.2 GiB | 16,384 | 30.5 GiB | On the device | 249 to 399 tokens a second |
| Qwen3-Next-80B-A3B-Thinking | Q2_K | 26.2 GiB | 32,768 | 31.0 GiB | On the device | 174 to 279 tokens a second |
453 more need KV cache numbers from a source before the check can size them.
How the estimate works
For each model, the check picks the largest quantization that fits, then the longest context that fits with it. A model too large for the graphics card can still run with some layers in computer memory, more slowly.
Memory kept for the system
- A graphics cardMay use all of its memory.
- Unified memoryKeeps about a quarter to a third for the system.
- Computer memory aloneKeeps a fifth for the system.
Enter your own memory to change these shares.
Where the numbers come from
Memory and bandwidth for each preset come from the vendor pages below, read September 24, 2026. Model sizes and architecture numbers come from each model's page. Add a bandwidth figure to any preset to see its speed.
- NVIDIAGeForce graphics cards comparedGeForce RTX 4090GeForce RTX 3090RTX PRO 6000 Blackwell WorkstationL40SA100H100DGX Spark
- AMDRyzen AI Max+ 395Radeon RX 7900 XTX
- AppleMacBook AirMacBook ProMac miniMac Studio
Paid links
Every listing on this page is unpaid. The order and contents of each list come from the named sources.