# naxiv

> Independent, hands-on guides, reviews and benchmarks for running LLMs and AI tools — on your own hardware or rented cloud GPUs. Every recommendation is tested first-hand; benchmark numbers are measured on real hardware (or rented GPUs), never estimated.

naxiv covers the full local-AI stack: budget single-board computers, used and new GPUs, mini-PCs and Macs, the software to run models (Ollama, llama.cpp, LM Studio), and cloud-GPU rental services. The site is authored by Pedro Santos, an embedded/AI/robotics engineer, and exists to help people run AI without vendor hype or guesswork.

## Guides & reviews

- [The cheapest way to run a local LLM](https://naxiv.com/posts/cheapest-way-to-run-local-llm/): budget ladder from an €80 Raspberry Pi to a used RTX 3090.
- [RTX 3090 vs RTX 4090 for local AI](https://naxiv.com/posts/rtx-3090-vs-4090-local-ai/): same 24 GB VRAM — when the 4090's speed is worth the premium.
- [Renting a cloud GPU: RunPod vs Vast.ai](https://naxiv.com/posts/renting-cloud-gpu-runpod-vs-vastai/): run big models by the hour without owning hardware.
- [How much VRAM do you need to run LLMs?](https://naxiv.com/posts/how-much-vram-to-run-llms/): rule of thumb + exact VRAM table from 3B to 70B at 4-bit.
- [Ollama vs llama.cpp vs LM Studio](https://naxiv.com/posts/ollama-vs-llamacpp-vs-lmstudio/): which local-AI tool to use and why.
- [Run an LLM on a Raspberry Pi 5](https://naxiv.com/posts/run-llm-on-raspberry-pi-5/): hands-on tiny-model inference on an SBC.

## Browse

- [All articles](https://naxiv.com/posts/): full index of guides, reviews and benchmarks.
- [Benchmarks](https://naxiv.com/benchmarks/): measured tokens/sec per (model × GPU × quant).
- [About & how we test](https://naxiv.com/about/): author background and testing methodology.

## Optional

- [Tags](https://naxiv.com/tags/): browse articles by topic.
- [RSS feed](https://naxiv.com/rss.xml): subscribe to new articles.