A lookup table for the question that precedes every local-AI hardware purchase. Figures are weight-only unless noted, computed from total parameter counts, and rounded. Add cache and overhead before deciding anything.
The Arithmetic
Memory for weights is parameter count multiplied by bytes per parameter: 2 bytes at BF16, 1 at FP8 or Q8, roughly 0.5 at Q4 or MXFP4. For sparse mixture-of-experts models use total parameters, not active. Then add key-value cache, which scales with context length and can exceed the weights themselves at very long context, plus 10 to 15 percent runtime overhead.
Memory by Model
| Model | Total params | Q4 | Q8 | BF16 | Minimum practical hardware at Q4 |
|---|---|---|---|---|---|
| Small models (3B class) | 3B | ~1.5 GB | ~3 GB | ~6 GB | Any modern laptop |
| 8B class | 8B | ~4 GB | ~8 GB | ~16 GB | 8GB GPU, 16GB Mac |
| Dense 32B class | 32B | ~16 GB | ~32 GB | ~64 GB | 24GB GPU, 32GB Mac |
| Nemotron TwoTower | ~60B | ~30 GB | ~60 GB | ~120 GB | RTX 5090 32GB, tight |
| Llama 4 70B | 70B | ~35 GB | ~70 GB | ~140 GB | 48GB VRAM or 64GB unified |
| Qwen3-Coder-Next | 80B | ~40 GB | ~80 GB | ~160 GB | 48GB VRAM or 64GB unified |
| gpt-oss 120B | 120B | ~60 GB | ~120 GB | ~240 GB | 128GB unified |
| MiniMax M3 | 428B | ~214 GB | ~428 GB | ~856 GB | 4x H100 80GB |
| GLM-5.2 | 753B | ~377 GB | ~753 GB | ~1,506 GB | 8x H100 80GB |
| Kimi K3 | 2.8T | ~1,400 GB | ~2,800 GB | ~5,600 GB | 16x H100 or 2x H200 node |
Key-Value Cache, the Part People Forget
Cache memory grows linearly with context length and is not small at 2026 context windows. Rough additional requirement for a 70B-class model:
| Context length | Approximate additional cache memory |
|---|---|
| 8K tokens | ~1-2 GB |
| 32K tokens | ~5-8 GB |
| 128K tokens | ~20-30 GB |
| 1M tokens | ~150-250 GB |
At a million tokens the cache dwarfs the weights. This is why million-token context windows are widely advertised and rarely used at full length outside data centres, and why cache-compression techniques matter as much as weight quantisation now.
Practical Sizing
Take the Q4 weight figure, add cache for the context length you will actually use rather than the maximum advertised, add 15 percent, and leave room for the operating system. On a 128GB machine running a 70B model at 32K context, that is roughly 35GB of weights plus 8GB of cache plus overhead, so around 50GB committed and plenty of headroom. On the same machine at 128K context it is closer to 75GB, which is still fine. At 1M it does not fit.
Brand Visibility Implications
The models in the top half of this table run on hardware individuals own, and those are the models generating unobservable brand answers today. The bottom half needs a data centre, which in practice means either a hosted API you could monitor or a large enterprise deployment you cannot. See the local-LLM visibility blind spot.
Methodology
Vendor specifications come from NVIDIA, Apple, and model-card publications. Throughput figures aggregate community benchmark reporting from the llama.cpp discussions, the MLX repository, and published independent test suites. Single-stream decode unless stated otherwise. Ranges rather than point values are used wherever independent runs disagree, which is most of the time: quantisation format, prompt length, thermal state, and runtime version each move these numbers by more than the differences being measured. Treat every figure as an order-of-magnitude guide, not a specification. Updated quarterly.
How Presenc AI Helps
Presenc AI measures how brands are represented across these models regardless of where they are served.