How much VRAM does this model need?
Use model type, parameter count and precision to estimate inference VRAM requirements and determine the GPU memory capacity you need.
1. Choose a model and parameter count
Language and multimodal understanding models. Estimate weights first, then add context memory where supported.
1B = one billion parameters. Use total parameters, including for MoE models, rather than the active parameter count.
2. Choose weight precision
Compare ideal weight memory side by side. 16-bit includes FP16 / BF16; 8-bit includes FP8 / INT8. Real quantization adds overhead, and the model and runtime must support the selected precision.
Included weights and advanced settings
Usually no changes are needed. Weights already included in a full checkpoint are not counted twice; required separate weights are summed automatically.
Refine the estimate: context and runtime reserve
For supported model architectures, set context length and simultaneous requests. Add other runtime overhead from measurements; it is not estimated automatically.
Automatic cache estimation is not available for this model. The result includes only weights and your runtime reserve; leave room for unestimated overhead.
Extra runtime reserve (GiB)
Use your own experience or measurements to reserve space for activations, workspace and loading. This is an assumed budget, not a measured prediction.
3. Check VRAM requirements
Approximate weight memory
16 GB
≈ 15 GiB
Weight memory is a starting point. Inference also needs room for caches, activations and runtime overhead. This estimate assumes all selected weights stay in VRAM; frameworks that support offloading can reduce VRAM use.
View weight breakdown
- Custom component
- 16 GB
VRAM capacity reference
Comparison basis: Weights only
Compare capacities with weights and any added inference budget. Spare capacity only means the selected budget is below capacity; it does not guarantee execution.
Compare with available VRAM or one unified-memory pool. Do not add a discrete GPU's VRAM to system RAM.
The current budget is below this capacity; uncounted inference overhead still needs roomDifference: 9 GiB
How much space do model weights take?
A parameter is a number stored in a model. Its storage cost depends on precision. A pipeline can require several models, so count each required component once.
Ideal weights (GB) = parameters (B) × bytes per parameter
7B × 2 bytes (BF16) ≈ 14 GB ≈ 13 GiB
1B = one billion parameters; GB = 10⁹ bytes; GiB = 2³⁰ bytes. We show both units to keep file sizes and memory capacities clear.
| Precision | Bytes / parameter | Ideal 7B weights |
|---|---|---|
| FP32 | 4 | 28 GB |
| BF16 / FP16 | 2 | 14 GB |
| FP8 / INT8 | 1 | 7 GB |
| INT4 | 0.5 | 4 GB |
INT4 is an ideal 0.5-byte estimate. Real quantized formats use scales, metadata and mixed precision. FP8 and INT8 have different numeric behavior; neither availability nor quality is guaranteed by this table.
Where do these weights live?
With GPU-resident inference, loaded weights usually occupy VRAM. The total here is a baseline, not a test of whether your hardware can run the model.
Offloading can keep some components or layers in system RAM and move them to the GPU as needed. It can lower peak VRAM, but may cost speed and needs framework support. Adding RAM and VRAM does not produce one interchangeable pool.
Unified-memory devices share one pool between CPU and GPU. Count it once, after reserving space for the system and other apps. Multi-GPU splitting is outside this tool.
Why does inference need more?
Chat generation often keeps a KV cache. Longer context and more simultaneous requests increase it, but architectures such as hybrid attention and MLA need their own formulas.
Image/video generation needs activations and temporary space. Higher resolution, more frames and larger batches can make these substantial even when weight size stays unchanged.
Loading, conversion and decompression can temporarily require more RAM or VRAM. Component offloading can make the sum of all files differ greatly from peak VRAM.
Preset sources and assumptions
Checked 2026-10-11. Parameter presets use published counts; file presets use selected artifacts at pinned revisions. Alternatives and duplicate pipeline layouts are not summed. Counts and file sizes are references, not runtime measurements.
- Qwen3.8-27B · BF16
Checked: 2026-10-11
- Qwen3-4B · BF16
Checked: 2026-10-11
- Qwen3-8B · BF16
Checked: 2026-10-11
- Qwen3-32B · BF16
Checked: 2026-10-11
- Kimi-K2.5 · native INT4
Checked: 2026-10-11
- Qwen-Image-2.1
Checked: 2026-10-11
- Stable Diffusion XL · FP16
Checked: 2026-10-11
- FLUX.1-schnell
Checked: 2026-10-11
- MiniMax H3 · FL2VA / full weights
Checked: 2026-10-11
- Wan2.1 T2V · 1.3B
Checked: 2026-10-11
- EmbeddingGemma 2 · multimodal
Checked: 2026-10-11
- EmbeddingGemma 2 · text only
Checked: 2026-10-11
- Qwen3-Embedding-4B · BF16
Checked: 2026-10-11
- MiniMax H3 · Q2_K
Checked: 2026-10-11
- MiniMax H3 · Q4_K
Checked: 2026-10-11
- Qwen3-8B · Q4_K_M
Checked: 2026-10-11
Calculations run in your browser. Presets require no live Hugging Face request or weight download. Source links open the public model repositories.