Guides
Practical explanations backed by the same data as our calculators. Every number in these guides is calculated, not copied.
Start here
How Much VRAM Do You Need to Run an LLM?A plain-English guide to LLM memory: the rule of thumb, what weights and KV cache cost, how much VRAM 7B, 14B,…Best Local LLMs for Every VRAM SizeThe best open models to run locally in 2026, grouped by how much VRAM you have: 8, 12, 16, 24, 32, 48 and 96+ …Q4 vs Q5 vs Q8: Quantization ExplainedWhat Q4_K_M, Q5_K_M, Q6_K and Q8_0 mean, how much memory each saves, how much quality you lose, and how GGUF c…KV Cache ExplainedWhat the KV cache is, the formula for its size, why some models need 16 GB for 128K context and others under 1…
Choosing hardware
Best GPU for Local LLMsWhich GPU to buy for running LLMs locally, by budget and model size: RTX 3060 12GB, used RTX 3090, RTX 5090, R…Best Mac for Local LLMsHow much Mac memory you need for local LLMs, how much the GPU can actually use, M4 vs M5 speed, and which mode…Running Local LLMs on AMD GPUsWhich AMD Radeon GPUs work for local LLMs, ROCm vs Vulkan, Ollama and LM Studio support, how fast they are com…Multi-GPU Setups for Local LLMsCan you combine VRAM across GPUs? How llama.cpp, Ollama, vLLM and ExLlama split models, layer split vs tensor …How to Run a 70B Model on Consumer GPUsEvery practical way to run Llama 3.3 70B and other 70B models at home: two RTX 3090s, a 48 GB card, Apple Sili…
Software and settings
Ollama VRAM Settings That Actually MatterHow to control Ollama's GPU memory: set num_ctx, enable flash attention, quantize the KV cache, limit parallel…LM Studio GPU Offload ExplainedWhat LM Studio's GPU offload slider does, how many layers to offload for your VRAM, the context length and KV …CPU Offloading: Running Models Bigger Than Your VRAMWhat happens when a model doesn't fit in VRAM, how much speed you lose with partial CPU offload, DDR4 vs DDR5,…Why Does My Model Use More VRAM Than Expected?Fix CUDA out of memory errors and unexpected VRAM use in Ollama, llama.cpp, LM Studio and vLLM: context defaul…
Models and use cases
Mixture-of-Experts Models and VRAMWhy mixture-of-experts models like gpt-oss, Qwen 3.6 35B-A3B and DeepSeek need memory for all parameters but r…Best Local Coding Models for a 24 GB GPUWhich coding models to run on an RTX 3090, 4090 or other 24 GB GPU, with the context each can hold for coding …How Much VRAM Do You Need to Fine-Tune an LLM?VRAM needed to fine-tune 3B to 70B models with QLoRA, LoRA and full fine-tuning, what uses the memory, and set…
Images and video
Running Flux and Other Image Models on Low VRAMHow much VRAM Flux needs in FP16, FP8 and GGUF, and how to run FLUX.1 dev, FLUX.1 Kontext, FLUX.2 and Z-Image …How Much VRAM Do You Need for AI Video?VRAM needed for local AI video models (Wan 2.2 14B and 5B, LTX-2, HunyuanVideo) by resolution, frame count and…