Free tool
LLM VRAM calculator
Pick a model, quantization, and rig. For a unified-memory machine, enter its installed memory. We estimate whether it fits in your VRAM or unified memory (weights + KV cache + overhead). Compare the estimate with submitted speed measurements for the model and hardware family where available; the run's quantization, context and runtime may differ.
Measured memory, not estimates. On one RTX 5090 (32 GB) with llama.cpp 9725a31 and a 16,384-token context configured, Qwen3.8-27B Q4_K_M (18.97 GB file) loaded all 65/65 layers on the GPU and the runtime reported 18.62 GiB allocated for model, context and compute buffers (runs r_c3ps9wjygi9, r_dz5a3lzxhrj, r_7mod38qsldj). Gemma 4 12B IT QAT Q4_0 (6.98 GB file) loaded 49/49 layers with 7.75 GiB allocated (runs r_c2podhe9qwc, r_8xmm65n1fhq, r_v5arqmazf31). These are runtime-log allocations, not peak device memory; the calculator above is an estimate for other models and contexts.
Fit result
Qwen2.5-14B-Instruct · Q4_K_M (4-bit) · 4K context · M3 Pro
Estimated fit at maximum memory
Estimated need: ~10 GB of the M3 Pro's 29 GB usable unified memory — ~19 GB to spare.
This estimate assumes the family’s maximum 36 GB configuration, with about 20% reserved for the OS. Enter your installed memory above to estimate a smaller configuration. Available memory and runtime limits can differ; this is not a measured fit test.
Weights
8.4 GB
Q4_K_M (4-bit)
KV cache
1.1 GB
4K ctx, fp16
Overhead
0.8 GB
runtime
And how fast?
signed a signed run measured 14.41tok/s decode on this model and hardware family; memory, quantization and runtime may differ from this estimate — see the run.
How the estimate works
Memory ≈ weights + FP16 KV cache + a fixed runtime allowance. Source-backed model profiles use attention dimensions and exact file sizes where shown; other selections use parameter-count and quantization heuristics. A file size is a planning input, not resident VRAM. All values are decimal GB and 1K means 1,000 tokens. Context includes input and generated tokens for one sequence. For mixture-of-experts models the total parameter count drives memory — all experts must be resident — even though only a few billion are active per token (which is why they decode fast). Apple unified memory reserves ~20% for the OS. These are estimates, not guaranteed upper bounds. Recurrent state, runtime buffers, tensor precision and vision components can change memory use. Allow extra memory for other applications and simultaneous requests. The submitted tok/s describes the linked run, not a speed guarantee for your selected settings. Read the full methodology.
VRAM questions, answered from measured runs
How much VRAM does a 27B model need?
Measured: Qwen3.8-27B at Q4_K_M used 18.62 GiB of runtime GPU allocation on an RTX 5090 with a 16,384-token context configured (all 65 layers on GPU). Longer contexts need more; use the calculator for your context length.
Will Qwen3.8-27B fit on a 24 GB GPU?
We have not measured it on a 24 GB card yet. The measured 18.62 GiB allocation at 16K context on a 32 GB RTX 5090 suggests it can load, but context headroom on 24 GB is untested. Check your context in the calculator.
How much VRAM does Gemma 4 12B need?
Measured: Gemma 4 12B IT QAT Q4_0 used 7.75 GiB of runtime GPU allocation (6.98 GB file, all 49 layers on GPU) on an RTX 5090 with a 16,384-token context configured.
What uses VRAM besides the model file?
The KV cache (grows with context length), compute buffers and runtime overhead. In our measured Qwen3.8-27B run the 18.97 GB file became 18.62 GiB of allocation including context and buffers at 16K configured context.
Is this calculator's number a measurement?
No. The calculator estimates weights + KV cache + overhead. Measured allocations are shown only where we have signed runs, each linked to its run IDs.