Halkali merkez neighborhood fatih st ozgur apt no 46 , Kucukcekmece , Istanbul , 34303 , TR
This tool goes beyond multiplying parameter count by bit width. It separates model weights, approximate GGUF bits-per-weight, KV cache for standard attention, window behavior for hybrid attention, runtime headroom and usable hardware memory. Results are planning estimates; actual use varies by runtime, driver, backend, batch and model file.
This is a capacity-planning estimate; multimodal projectors, speculative decoding, concurrent users, large batches, CUDA graphs and runtime workspaces may consume additional memory.
Model weight is estimated from total parameters × effective quantization bits-per-weight / 8. GGUF K-quants keep some tensors at different precisions, so formats such as Q4_K_M are not exactly 4.0 bits.
For standard transformer attention, KV cache is approximated as 2 × layers × KV heads × head dimension × tokens × bytes per KV element. GQA can reduce cache because KV heads are fewer than attention heads.
For local/global attention such as Gemma 3, local layers cache only the local window. For GatedDeltaNet + full-attention hybrids such as Qwen3.8, the calculator directly models the context-growing full-attention KV portion and treats additional GatedDeltaNet state conservatively inside runtime headroom.
| Model | Parameters | Architecture | Context | KV profile |
|---|---|---|---|---|
| Qwen3 8B | 8.2B | 36 layer · GQA | 40K | 8 KV heads · 128 dim |
| Qwen3.8 27B | 27B | 64 layer · GatedDeltaNet + Full Attention | 262K | Every 4th layer full attention |
| Qwen2.5 / DeepSeek R1 Distill 32B | 32.5B | 64 layer · GQA | 32K+ | 8 KV heads · 128 dim |
| Llama 3.3 70B | 70.6B | 80 layer · GQA | 128K | 8 KV heads · 128 dim |
| Gemma 3 27B | 27B | Local/Global Attention | 128K | Local window aware |
| Mistral Small 3.1 24B | 24B | Dense Transformer | 128K | Planning profile |
| DeepSeek V3 / R1 Full | 671B | MoE + MLA | 128K | Runtime-dependent estimate |
K-quants include block scales and may keep selected tensors at higher precision, so effective bits-per-weight is above 4. The calculator uses an approximate 4.85 bpw for Q4_K_M.
Q4-class weights for many 32B models approach roughly 20 GB. Context and runtime memory can make a 24 GB card tight, often requiring shorter context or CPU offload.
Capacity can increase with tensor/model parallel or layer splitting, but runtimes do not always expose multiple GPUs as one perfect memory pool. The calculator totals capacity and applies safety headroom.
Unified memory is shared by CPU and GPU, so the whole amount cannot safely be dedicated to the model. Apple profiles keep a larger system reserve.
Autoregressive generation retains key/value states for previous tokens. In standard attention, KV cache grows approximately linearly with context.
DeepSeek V3 uses Multi-Head Latent Attention rather than a standard KV layout. Weight memory can be estimated, while context memory depends more strongly on implementation and runtime.
After estimating VRAM, choose GPU capacity with enough headroom. Long context, concurrency and large batch sizes require more than simply fitting the model weights.
Turkey (Turkish)
Germany (German)
Worldwide (English)
