Open Dataset • CC BY 4.0

Local LLM VRAM & KV-Cache Scaling Dataset (2026 Lab Matrix)

Empirical memory profiling on our test bench across NVIDIA RTX 4090 (24GB), RTX 4070 Ti Super (16GB), and RTX 4060 (8GB). Measured under CUDA 12.6, Ollama v0.5.8, and llama.cpp b4120.

How to Cite This Benchmark

BibTeX / APA available
PraveenTechWorld Research Lab. (2026). Local LLM VRAM & KV-Cache Benchmark Matrix (v1.2).
https://www.praveentechworld.com/benchmarks/vram-llm

1. Model Weight VRAM Footprint by Quantization

ModelParametersFP16 WeightQ8_0 WeightQ4_K_M WeightMin GPU Recommended
DeepSeek-R1-Distill-Qwen-7B7.6B15.2 GB8.1 GB4.7 GBRTX 3060 / 4060 (8GB)
DeepSeek-R1-Distill-Qwen-14B14.8B29.6 GB15.6 GB9.0 GBRTX 4060 Ti (16GB)
Phi-4 (Microsoft)14.7B29.4 GB15.4 GB8.9 GBRTX 4060 Ti / 4070 (16GB)
Llama-3.3-70B-Instruct70.6B141.2 GB74.2 GB42.8 GBDual RTX 3090 / 4090 (48GB)

2. KV-Cache Memory Scaling Overhead by Context Depth

VRAM is consumed not only by model weights, but by active attention key-value caches. Here is the measured overhead in gigabytes added to base model weights:

Model Architecture4K Context8K Context16K Context32K Context
Qwen 2.5 / DeepSeek-R1 (7B)0.32 GB0.64 GB1.28 GB2.56 GB
Phi-4 (14B)0.45 GB0.90 GB1.80 GB3.60 GB
Llama 3.3 (70B with GQA)0.62 GB1.25 GB2.50 GB5.00 GB