LLM VRAM requirements, model by model
Every model page shows the computed weight sizes at Q4 and Q8 and a verdict matrix against common hardware. Numbers are estimates from parameter counts — the formula is on every page.
Llama 3.1 8B
Meta · 8B · 2024-07
Meta’s workhorse 8B model. The default first local LLM — strong general performance at a size almost any modern GPU can host.
Q4 ≈ 4.6 GB · Q8 ≈ 8.5 GB
Llama 3.3 70B
Meta · 70B · 2024-12
Meta’s flagship open model. Near-frontier quality, but needs roughly 40 GB at Q4 — multi-GPU, a 24 GB card at aggressive quantization, or a big-memory Mac.
Q4 ≈ 40.6 GB · Q8 ≈ 74.2 GB
Mistral 7B
Mistral AI · 7B · 2023-09
The model that made 7B respectable. Compact, fast, and a long-standing favorite for CPU-only and 8 GB GPU setups.
Q4 ≈ 4.1 GB · Q8 ≈ 7.4 GB
Mixtral 8x7B
MoEMistral AI · 47B · 2023-12
A sparse mixture-of-experts model: 47B total parameters, ~13B active per token. Fast for its quality — but all 47B of weights still must fit in memory.
Q4 ≈ 27.3 GB · Q8 ≈ 49.8 GB
Qwen2.5 7B
Alibaba · 7B · 2024-09
Alibaba’s 7B, strong on coding and multilingual tasks. A top pick in the 8 GB GPU class.
Q4 ≈ 4.1 GB · Q8 ≈ 7.4 GB
Qwen2.5 14B
Alibaba · 14B · 2024-09
The 14B tier of Qwen2.5 — a clear quality step from 7B that still fits a 12 GB card at Q4.
Q4 ≈ 8.1 GB · Q8 ≈ 14.8 GB
Qwen2.5 32B
Alibaba · 32B · 2024-09
The sweet spot of the Qwen2.5 family for many: flagship-adjacent quality that fits a 24 GB card at Q4.
Q4 ≈ 18.6 GB · Q8 ≈ 33.9 GB
Qwen2.5 72B
Alibaba · 72B · 2024-09
Qwen2.5 at near-frontier scale. Comparable memory class to Llama 3.3 70B — around 43 GB at Q4.
Q4 ≈ 41.8 GB · Q8 ≈ 76.3 GB
Gemma 2 9B
Google · 9B · 2024-06
Google’s 9B model, known for punching above its size in conversational quality. Fits comfortably on 8–12 GB cards at Q4.
Q4 ≈ 5.2 GB · Q8 ≈ 9.5 GB
Gemma 2 27B
Google · 27B · 2024-06
The large Gemma 2. Around 16 GB at Q4 — 16 GB-class cards and 32 GB Macs are the natural hosts.
Q4 ≈ 15.7 GB · Q8 ≈ 28.6 GB
Phi-4 14B
Microsoft · 14B · 2024-12
Microsoft’s small-model line focused on reasoning quality per parameter. Efficient at 14B and friendly to 12 GB cards.
Q4 ≈ 8.1 GB · Q8 ≈ 14.8 GB
Muse Spark Open 816B
MoEMeta · 816B · 2026-10
Meta’s open-weight flagship (816B total, 42B active, 1M-token context). Runs on two Mac Studio-class machines at inference providers — consumer GPUs need not apply. Included here as the honest upper bound of the chart.
Q4 ≈ 473.3 GB · Q8 ≈ 865 GB