Choose a model and weight precision, add operating headroom, and compare the indexed single-device options that clear the resulting memory target. Results are planning estimates—not runtime support or live-price guarantees.
Smallest memory footprint. Headroom is a user-controlled planning buffer; actual use varies by runtime, context length, KV cache, concurrency, and offloading.
Source: Official Qwen GGUF model card ↗The calculator begins with GPU Hunter's model-specific memory baseline for Q4, Q8, or FP16 weights. The optional headroom setting increases that baseline before comparing it with the memory capacity of every device in the current index.
This is intentionally a capacity screen, not a promise that a model will run in every configuration. Context length, KV cache, batching, concurrency, runtime allocations, drivers, and offload strategy all affect real memory use. Validate the final setup in the runtime you plan to use.
Lower-bit quantization compresses model weights. That can make a model viable on less expensive hardware, but formats differ in quality, kernel support, and speed. Treat Q4, Q8, and FP16 as distinct deployment choices rather than interchangeable labels.
These values are GPU Hunter's weight-size baselines before optional operating headroom. They are not measurements of every possible context length or runtime.
| Model | Q4 | Q8 | FP16 | Listed context | Source |
|---|---|---|---|---|---|
| Qwen3 32B | 19GB | 36GB | 64GB | 128K | Model card ↗ |
| Qwen2.5 72B | 44GB | 77.5GB | 145GB | 128K | Model card ↗ |
| Qwen3 235B-A22B | 132GB | 240GB | 470GB | 128K | Model card ↗ |
| Llama 3.3 70B | 40GB | 75GB | 140GB | 128K | Model card ↗ |
| DeepSeek V3 | 380GB | 700GB | 1300GB | 128K | Model card ↗ |
| Llama 3.1 8B | 5GB | 9GB | 16GB | 128K | Model card ↗ |
| Qwen3 14B | 8GB | 15GB | 28GB | 128K | Model card ↗ |
| Mistral 7B | 4GB | 8GB | 14GB | 32K | Model card ↗ |
| Gemma 2 27B | 16GB | 30GB | 54GB | 8K | Model card ↗ |
| Codestral 22B | 13GB | 24GB | 44GB | 32K | Model card ↗ |
It depends on the model size, weight precision, context length, runtime overhead, KV cache, and concurrency. Select a known model and quantization to get a planning baseline, then add headroom for the way you intend to run it.
Q4 stores weights at roughly four-bit precision and generally uses the least memory. Q8 uses more memory in exchange for higher weight precision. FP16 is a full-weight 16-bit baseline and usually requires substantially more memory. Exact quality and performance depend on the quantization method and runtime.
Fifteen percent is GPU Hunter's default planning allowance, not a universal requirement. Longer context, larger KV caches, batching, concurrent users, and runtime behavior may require more. CPU or multi-GPU offloading changes the calculation.
No. Enough memory only clears the capacity constraint. Confirm that your runtime, drivers, model format, kernels, operating system, and device architecture are supported.
No. They are dated reference values from the GPU Hunter dataset, last updated Apr 30, 2026. Always verify the current price, condition, seller, and availability.