Free planner

VRAM / GGUF fit estimator

ESTIMATE · last verified 2026-09-11

Rough fit check before you pull a multi-GB GGUF. Numbers stay in this tab — no upload.

GPU VRAM (GB) System RAM free (GB) optional — used for offload note Model params (B)

7B 8B 13B 14B 32B 70B

Quantization Q2_K (~0.3 bpp ESTIMATE) Q3_K_M (~0.4 bpp ESTIMATE) Q4_K_M (~0.5 bpp ESTIMATE) Q5_K_M (~0.6 bpp ESTIMATE) Q6_K (~0.7 bpp ESTIMATE) Q8_0 (~1.0 bpp ESTIMATE) FP16 (~2.0 bpp ESTIMATE) Known file size (GB) optional — overrides params × bpp Context tokens

4k 8k 16k 32k

UI on same GPU (Open WebUI etc.) — +1 GB ESTIMATE headroom Embeddings on GPU — +1 GB ESTIMATE headroom

ESTIMATE bytes/param table

Planning multipliers only — real GGUF files vary by architecture and tensor layout. Prefer a known file size when you have one.

Quant	Bytes/param (ESTIMATE)
Q2_K	~0.3
Q3_K_M	~0.4
Q4_K_M	~0.5
Q5_K_M	~0.6
Q6_K	~0.7
Q8_0	~1.0
FP16	~2.0

Formulas (ESTIMATE): Weights GB ≈ params_B × bpp (or known file size). Crude ctx tax ≈ 0.5 GB base + (tokens / 8192) × 0.5 GB. UI / embeddings toggles add +1 GB each.

Fit rules: Full GPU if total < VRAM × 0.85; Likely offload if total < VRAM × 0.85 + RAM_free × 0.5; else Probably won't fit comfortably.

Honest limits

• Apple unified memory is not the same as discrete NVIDIA VRAM bands.
• This is not tokens/sec and not a promise the card will feel fast.
• Exact GGUF file size on the model card always wins over the formula.
• Don't plan to the last gigabyte — leave headroom for the runtime.

Related

• VRAM & RAM for local LLMs — editorial bands
• Storage for local models — disk / NVMe notes
• Context memory helper — leftover VRAM vs tokens
• Free planners hub

Need a printable worksheet? Free sample · packs $39 / $29.

← Free planners · Tools