Research

How Much RAM to Run Each Open-Weight Model

Memory requirement lookup table for the major 2026 open-weight models at Q4, Q8, and BF16, including key-value cache overhead at long context and the minimum hardware that clears each tier.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: July 2026

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

ModelTotal paramsQ4Q8BF16Minimum practical hardware at Q4
Small models (3B class)3B~1.5 GB~3 GB~6 GBAny modern laptop
8B class8B~4 GB~8 GB~16 GB8GB GPU, 16GB Mac
Dense 32B class32B~16 GB~32 GB~64 GB24GB GPU, 32GB Mac
Nemotron TwoTower~60B~30 GB~60 GB~120 GBRTX 5090 32GB, tight
Llama 4 70B70B~35 GB~70 GB~140 GB48GB VRAM or 64GB unified
Qwen3-Coder-Next80B~40 GB~80 GB~160 GB48GB VRAM or 64GB unified
gpt-oss 120B120B~60 GB~120 GB~240 GB128GB unified
MiniMax M3428B~214 GB~428 GB~856 GB4x H100 80GB
GLM-5.2753B~377 GB~753 GB~1,506 GB8x H100 80GB
Kimi K32.8T~1,400 GB~2,800 GB~5,600 GB16x 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 lengthApproximate 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.

Frequently Asked Questions

Roughly 35GB for weights at Q4 quantisation, plus key-value cache of about 5-8GB at 32K context or 20-30GB at 128K, plus around 15 percent runtime overhead. In practice 48GB of VRAM or 64GB of unified memory is the sensible minimum, and 128GB is comfortable at long context.
Multiply total parameters by bytes per parameter: 0.5 at Q4, 1 at Q8, 2 at BF16. For sparse mixture-of-experts models use total parameters, not active. Then add key-value cache, which scales with context length, plus 10 to 15 percent runtime overhead.
Substantially. For a 70B-class model, key-value cache adds roughly 1-2GB at 8K tokens, 20-30GB at 128K, and 150-250GB at 1M. At the longest context windows the cache exceeds the weights, which is why million-token contexts are rarely used at full length outside data centres.
Roughly a 220B-parameter model at Q4 on weights alone, though 120B-class is the comfortable working ceiling once cache and overhead are included. That covers gpt-oss 120B and everything below it, but not MiniMax M3 at 428B or GLM-5.2 at 753B.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.