Find the perfect AI model
for your computer

It depends on your GPU's VRAM and memory bandwidth. As a rule of thumb, an 8 GB card runs 7–8B models at Q4, 12 GB runs up to ~14B, 16–24 GB runs 24–32B, and 48 GB+ (or Apple Silicon unified memory) runs 70B models. The Local AI Advisor detects your exact hardware and lists every model that fits, how fast it runs, and the best quantization to use.

A rough rule: model VRAM ≈ parameters (in billions) × bits-per-weight ÷ 8, plus the KV cache for your context and ~1 GB of overhead. An 8B model at Q4_K_M needs about 4.8 GB of weights, so ~6–7 GB total — comfortable on an 8 GB card. A 70B model at Q4 needs roughly 40 GB. The advisor computes the exact footprint for any model, quant and context length.

A 70B model at Q4_K_M needs roughly 40 GB of memory, so a 48 GB card (like an RTX A6000), two 24 GB GPUs, or Apple Silicon with 64 GB+ of unified memory can run it fully in memory. On a 24 GB card it only runs by offloading part of the model to system RAM, which drops speed sharply — a 24–32B model is the sweet spot for a single 24 GB GPU.

Single-stream text generation is memory-bandwidth bound: each token reads the active weights once, so tokens/sec ≈ usable memory bandwidth ÷ active-weight bytes. For example an RTX 4090 (~1008 GB/s) runs an 8B model at Q4 around 120+ tokens/sec — far faster than reading speed. The advisor estimates realistic tokens/sec for every GPU, model and quantization.

No — you can run smaller models on the CPU alone, but it is much slower because system RAM bandwidth is far lower than GPU VRAM bandwidth. A dedicated GPU, or Apple Silicon with its fast unified memory, is strongly recommended for anything above ~3B parameters if you want a usable, interactive speed.

Q4_K_M is the default sweet spot — about 97.5% of full quality at ~4.8 bits per weight. Use Q5_K_M or Q6_K for near-lossless output if the model still fits in memory, and only drop to Q3, IQ3 or Q2 when you need to squeeze a larger model into limited VRAM.

Throughput is nearly identical — both run llama.cpp / GGUF under the hood, so speed is set by your hardware and quantization, not the app. Choose Ollama for CLI and automation, LM Studio for a friendly GUI (and its MLX backend on Apple Silicon), or vLLM / TabbyAPI when you need to serve many concurrent requests.

Yes. A laptop with a 6–8 GB dedicated GPU runs 7–8B models comfortably, and Apple Silicon MacBooks with 16 GB+ of unified memory run 8–14B models well (larger models with more memory). The advisor detects laptop GPUs and integrated graphics, and estimates the slower CPU-class speed when there is no dedicated GPU.

MoE models only activate a fraction of their parameters per token — Qwen3-30B-A3B activates about 3B of its 30B parameters. Because generation speed depends on active-weight bytes, it decodes roughly as fast as a 3B model while answering with 30B-scale quality, which is why it feels much quicker than a dense 32B model.