VRAM Calculator: How Much GPU Memory Do You Need?
Updated September 8, 2026. Model data captured from Hugging Face at pinned revisions; GPU capacities and price links come from our GPU specification database.
How much VRAM does your next AI model need? Pick a current model - DeepSeek V4-Pro, DeepSeek V4-Flash, GLM 5.3, GLM 5.3-Flash, Qwen 3.8-Flash-Next, Tencent Hy4-preview, GLM-4.5-Air and more - and the calculator estimates memory from the real checkpoint file sizes plus config-exact KV cache math. Then build your GPU pool: stack the cards you own or plan to buy, because most 2026 flagships do not fit on a single consumer GPU. Every Amazon-listed GPU carries a Check Price button; cards we could not verify on Amazon US (datacenter accelerators, RTX PRO 5000 Blackwell) intentionally have none.
Your GPU Pool
Add the GPUs you own or plan to buy. Flagship MoE models routinely need several cards - the verdict below stacks your pool capacity and reserves a small runtime overhead per card.
Single GPUs That Can Run This
Check Price buttons open Amazon through our affiliate link and only appear for cards we have verified as listed on Amazon US. Already know the model by name? The Model-to-GPU finder ranks tracked cards for popular open models with the full VRAM math shown.
Multi-GPU Pools & Alternatives
Show the exact formula trace
2 x attention layers x KV heads x head_dim x context x bytes x batch. Hybrid-attention models (Qwen 3.8-Flash-Next: 12 of 48 layers; GLM 5.3-Flash: 11 of 45) only grow KV on their attention layers; their linear-attention layers hold a small fixed state. Sparse-attention flagships (DeepSeek V4 CSA+HCA, GLM 5.3 DSA, Hy4 Gated DSA) are modeled as dense caches, so treat their KV line as an upper bound. Overhead adds runtime, ~5% framework buffers, workspace, and a 10% safety margin. DeepSeek V4 checkpoints ship FP4+FP8 mixed precision (experts FP4) - the artifact sizes reflect that.
How does the calculator estimate VRAM usage?
The estimate adds four components: the model weights (from exact checkpoint file sizes), the KV cache that grows with context length (from the architecture config), fixed overheads, and a safety margin. For MoE flagships the weight footprint follows the total parameter count, not the much smaller active count - every expert has to sit in memory even if only a few fire per token.
Two details matter for the 2026 generation. First, checkpoint precision varies by vendor: DeepSeek V4-Pro and V4-Flash ship FP4+FP8 mixed weights, GLM 5.3 and GLM 5.3-Flash ship native FP8, Hy4-preview and Qwen 3.8-Flash-Next ship BF16 - identical parameter counts therefore produce very different file sizes, and the calculator uses the real bytes. Second, hybrid and sparse attention changes the cache math: models like Qwen 3.8-Flash-Next keep full KV on only a fraction of their layers, and DSA/CSA-style flagships compress further at long context, so their calculated KV is deliberately labeled an upper bound.
The result is a planning estimate with honest labels: every number is either an exact file size, config-exact math, or clearly marked derived.
How do you use the VRAM calculator?
- Pick your model - current flagships and proven local models are grouped in the dropdown.
- Choose the checkpoint: official precision or a community GGUF quant, each labeled with its exact file size.
- Set context length (capped at the model's native window), KV cache precision, and batch size.
- Build your GPU pool: add one card or stack several. The verdict tells you if the configuration fits, is tight, or falls short - and by how much.
- Check the single-GPU list and multi-GPU pool suggestions, with Check Price buttons on every Amazon-listed card.
For anything serious, run the calculation twice: once at your target context and once a step longer. If the longer context crosses into the next capacity tier, that is the number that matters for your next purchase.
What fits in 8, 16, 24, and 32 GB of VRAM?
Capacity tiers map cleanly to workload classes: 8 GB runs small quantized models, 16 GB covers full-precision small models, 24 GB opens up mid-size territory, and 32 GB adds headroom for the largest quantized consumer workloads. Best for most local AI work: a 24 GB card. Best for headroom: 32 GB. For 2026 MoE flagships: a pool.
| Capacity | What comfortably fits | Example GPUs we track |
|---|---|---|
| 8 GB | 7B-8B models at INT4; light SDXL work | GeForce RTX 5060, Arc B580 territory |
| 16 GB | 7B-8B at FP16; 13-14B at INT4; comfortable SDXL | GeForce RTX 5080, RTX 4060 Ti 16GB, Radeon RX 9070 XT |
| 24 GB | 14B at FP16; ~30B at INT4; large SDXL batches | GeForce RTX 5090's smaller siblings: RTX 4090, RTX 3090, Radeon RX 7900 XTX |
| 32-96 GB | 70B at INT4; GLM-4.5-Air Q4 (~68 GiB); long contexts | GeForce RTX 5090, RTX PRO 6000 Blackwell (96 GB) |
| Pools / datacenter | DeepSeek V4-Flash (~145 GiB at Q4), GLM 5.3 / Hy4 (~435 GiB at Q4), DeepSeek V4-Pro (~805 GiB) | 4-8x RTX 5090 or RTX PRO 6000; H200 / MI300X / B200 class |
The pattern to internalize: quantization buys capacity, and pooling buys reach. A model that needs 24 GB at FP16 fits into far less at INT4 - and a flagship that outgrew your card entirely still runs split across two or four of them.
Why do some GPUs have no Check Price button?
Because we only show the button where we can verify a real Amazon US listing, re-checked through the Amazon Creators API. Datacenter accelerators (H100, H200, B200, B300, MI300X/MI325X/MI355X, A100) are sold through OEM and cloud channels, and the RTX PRO 5000 Blackwell had no legitimate US retail listing when we checked - those cards show no button rather than a dead or misleading link.
Can I split a model across two GPUs?
Yes, by splitting layers or tensors across cards - the calculator's GPU pool is built for exactly that. Two 24 GB cards hold a model that needs up to ~46 GB of weights plus cache, but the cards do not merge into one pool automatically: the runtime assigns layers to each card, and PCIe bandwidth becomes part of the cost. NVLink bridges close that gap on the few cards that support them. Prefer a single larger card when budget allows, and a pool when it does not.
Does anything else consume VRAM?
Yes: the KV cache, image-generation latents, and runtime overhead all sit alongside the weights. That is why the calculator's overhead lines exist. If you run a model and a tool such as ComfyUI together, budget for both.
How much extra VRAM should I leave free?
The calculator includes a 10% safety margin on weights plus KV. A tight fit today becomes an out-of-memory error after the next model update - if your pool verdict says "tight", size up.
Frequently Asked Questions
How accurate is the calculator's estimate?
Weights are exact file sizes and KV math is config-exact for standard attention. For sparse/hybrid-attention flagships the KV line is a labeled upper bound, and real peak memory can land lower. Buy against the computed number plus margin.
Is VRAM capacity or memory bandwidth more important?
Capacity decides what fits; bandwidth decides how fast it runs. A model that barely fits on a slow card still runs - slowly. Buy capacity first, then bandwidth.
Why does DeepSeek V4-Flash need less memory than its parameter count suggests?
The official checkpoint stores MoE expert weights in FP4 and the rest in FP8, so 284B parameters occupy ~149 GB instead of ~284 GB. The calculator uses the real artifact bytes.
Do image-generation models need the same VRAM as LLMs?
Similar logic, different mix: image models trade context length for large latent buffers, and batching multiplies them.
Sources
All model data captured September 8, 2026, from primary sources at pinned revisions:
- Hugging Face config.json + file-tree byte sums: deepseek-ai/DeepSeek-V4-Pro (@b5968e91), deepseek-ai/DeepSeek-V4-Flash (@60d8d707), zai-org/GLM-5.3 (@aca966e4), zai-org/GLM-5.3-Flash (@eb9eb208), tencent/Hy4-preview (@705d81ee), Qwen/Qwen3.8-Flash-Next (@de4b8e4d), zai-org/GLM-4.5-Air (@a24ceef6).
- Quantized artifact sizes: unsloth GGUF repos (GLM-5.3, GLM-5.3-Flash, GLM-4.5-Air, Qwen3.8-Flash-Next, DeepSeek-V4-Flash-0731) and AngelSlim/Hy4-preview-GGUF.
- Official parameter totals: DeepSeek V4 model card (284B/1.6T), z.ai GLM 5.3 release (743B/39B; Flash 320B/18B), Tencent Hy4-preview card (770B/49B), Qwen3.8-Flash-Next card (125B+51B n-gram+4B MTP).
- GPU capacities, categories, and Amazon availability: our GPU specification database (vram_gb spec + verified proxy links, 46 linked / 11 exempt).