# Local AI: Private Models on Your Own Hardware

Canonical: https://nosistech.com/local-ai/
Last updated: 2026-10-30
Author: Carlos, NosisTech LLC

Local AI means a language model that runs on a computer you own, not on a provider’s cloud. Your prompts and files stay on your machine, and local tasks have no per-token bill.

This page collects every NosisTech article on local AI. Each one documents a real build with the hardware, the commands, the measured speed, and the problems that came up.

## The Two Laptops Behind These Tests

| Laptop | Processor | GPU | System memory |
| --- | --- | --- | --- |
| Acer Predator Helios Neo 16S (PHN16S-71) | Intel Core Ultra 9 275HX | NVIDIA RTX 5070 Ti, 12GB GDDR7 | 32GB DDR5 at 6400 MT/s |
| ASUS TUF Dash F15 | Intel Core i7-11370H | NVIDIA RTX 3050 Ti, 4GB GDDR6 | Upgraded from the 8GB base model |

The 12GB laptop runs models up to 35 billion parameters, with part of each large model in system memory. The 4GB laptop runs a 2 billion parameter model, and a router sends hard prompts to a cloud model.

## Start Here

| Article | What it covers |
| --- | --- |
| [Running Local AI on a Gaming Laptop: Install and Setup](https://nosistech.com/running-local-ai-gaming-laptop-setup/) | The first install of Ollama and Open WebUI, from problems to fast settings. |
| [VRAM vs RAM in Local AI Models](https://nosistech.com/vram-vs-ram-local-ai-models-explained/) | Why a model that does not fit in GPU memory runs slower. |
| [Running Local AI on a $798 Laptop](https://nosistech.com/running-local-ai-on-a-798-laptop-the-nosistech-test/) | What a 4GB GPU can do across 71 real-world tasks. |

## Local AI on a Gaming Laptop

Nine articles on the 12GB laptop, in reading order. They move from memory basics to benchmarks, tool choice, and agent setup.

| # | Article | What it covers |
| --- | --- | --- |
| 1 | [VRAM vs RAM in Local AI Models](https://nosistech.com/vram-vs-ram-local-ai-models-explained/) | Qwen3 30B needs about 19GB, so roughly 9GB spills into slower system RAM. |
| 2 | [Running Local AI on a Gaming Laptop: Install and Setup](https://nosistech.com/running-local-ai-gaming-laptop-setup/) | Ollama and Open WebUI, three install problems, and tuned settings for Gemma4 26B and Qwen3 30B A3B. |
| 3 | [I Pushed My Local AI's Memory to the Limit](https://nosistech.com/local-ai-context-window-limit-testing/) | Context window tests from 4,096 to 32,768 tokens. |
| 4 | [Qwen3.6 35B-A3B on a 12GB Laptop GPU](https://nosistech.com/running-local-ai-gaming-laptop-setup-2/) | A 23GB model at 16 tokens per second, and why 2,048 tokens of context is too small. |
| 5 | [Qwen3, Sonnet, Opus, and Gemini Pro Compared](https://nosistech.com/running-local-ai-vs-cloud-giants/) | Where a local model holds up, and how I divide the work. |
| 6 | [LM Studio vs Ollama on a 12GB Laptop GPU](https://nosistech.com/lm-studio-vs-ollama-rtx-5070-ti-benchmark/) | One model and prompt: 28.00 tokens per second in LM Studio, 17.65 in Ollama. |
| 7 | [LM Studio vs Ollama for an Always-On AI Server](https://nosistech.com/lm-studio-vs-ollama-always-on-local-ai/) | Why speed was the wrong measure for a server that must survive a reboot. |
| 8 | [Maxing Out Gemma 4 on a Laptop GPU](https://nosistech.com/gemma-4-12b-local-optimization-mtp-hermes-agent/) | Gemma 4 12B from 39 to 114 tokens per second with Multi-Token Prediction and llama.cpp. |
| 9 | [Why Hermes Agent Rejects a 64k Context](https://nosistech.com/hermes-agent-64k-context-parallel-slots-fix/) | The context is split between parallel slots in llama-server, and one slot fixes it. |

## The $798 Laptop Test

Three articles test one question: can a 2021 gaming laptop with a 4GB GPU run useful private AI?

| # | Article | What it covers |
| --- | --- | --- |
| 1 | [Running Local AI on a $798 Laptop: The NosisTech Test](https://nosistech.com/running-local-ai-on-a-798-laptop-the-nosistech-test/) | Qwen3.5 2B in about 3GB of VRAM, tested with 71 real-world tasks. |
| 2 | [4GB GPU Made Laptop an AI Workstation](https://nosistech.com/4gb-gpu-made-laptop-an-ai-workstation-smart-cloud-routing/) | A router keeps everyday prompts local and sends hard ones to the cloud, tested with 40 scenarios. |
| 3 | [Your Own Private AI on a $798 Laptop: 91% Pass](https://nosistech.com/local-ai-router-91-percent-pass-798-laptop/) | 35 router tests, with 89% of queries answered on the laptop at no cost. |

## Home AI Lab and Remote Access

These articles turn the laptop into a small server that you can reach from a phone, and then lock it down.

| # | Article | What it covers |
| --- | --- | --- |
| 1 | [I Made My Local AI Accessible From Anywhere](https://nosistech.com/local-ai-accessible-anywhere-cloudflare-tunnel/) | Cloudflare Tunnel gives the laptop a permanent address on the free plan. |
| 2 | [Self-Hosted AI Lab Accessible From Anywhere](https://nosistech.com/ollama-open-webui-cloudflare-tunnel-windows-setup/) | A full Windows guide with NSSM, auto-start at boot, and a reboot test. |
| 3 | [Securing a Self-Hosted AI Lab with Cloudflare Zero Trust](https://nosistech.com/securing-self-hosted-ai-lab-cloudflare-zero-trust/) | Automated probes arrived within hours, and Zero Trust Access now checks every visitor. |

## A Private Voice Assistant

| # | Article | What it covers |
| --- | --- | --- |
| 1 | [A Private AI Voice Assistant on a Laptop](https://nosistech.com/jarvis-private-ai-voice-assistant-ollama-open-webui-windows/) | Whisper for speech, gemma3:12b for answers, and Kokoro for the voice. |
| 2 | [AI Assistant Tuning: What Broke and Finally Worked](https://nosistech.com/i-voice-tuning-modelfile-ollama-gemma-open-webui/) | The system prompt only held after it moved into an Ollama Modelfile. |

## Cloud Models, VPS Servers, and Cost

Local models handle much of the work. These articles cover the rest: a VPS server, cloud models behind one proxy, and the cost of each call.

| Article | What it covers |
| --- | --- |
| [I Built My Own AI Server on a VPS](https://nosistech.com/ai-server-hostinger-vps-litellm-gemini-open-webui/) | Ollama on a Hostinger VPS, with LiteLLM so Claude and Gemini sit beside local models. |
| [Connecting DeepSeek V4-Pro to a Hostinger VPS](https://nosistech.com/deepseek-v4-pro-hostinger-vps-litellm-open-webu/) | The LiteLLM and Open WebUI setup, and the price per million tokens. |
| [Three Hidden Patterns That Inflate Your AI API Bill](https://nosistech.com/three-hidden-patterns-that-inflate-your-ai-api-bill/) | Expensive default models, silent retries, and large models on small tasks. |

## Automations With n8n

| # | Article | What it covers |
| --- | --- | --- |
| 1 | [A Local AI Discount Monitor With n8n, Ollama, and Telegram](https://nosistech.com/n8n-ollama-gcp-cert-discount-monitor/) | A daily check of the Google Cloud certifications blog. |
| 2 | [From Laptop to Cloud: The Discount Monitor Moves to a VPS](https://nosistech.com/gcp-discount-monitor-vps-migration/) | Why a laptop that sleeps cannot run a daily job. |
| 3 | [A Google Search Console Automation With n8n](https://nosistech.com/n8n-google-search-console-automation/) | A daily sitemap check with a Telegram confirmation. |

## More From NosisTech

| Page | What it covers |
| --- | --- |
| [AI Security](https://nosistech.com/ai-security/) | Bunker, website hardening, and AI safety builds. |
| [AI Governance](https://nosistech.com/ai-governance/) | Free policy templates and governance builds. |
| [50+ Agent Patterns & Builds](https://nosistech.com/50-agent-patterns-builds/) | The full library of documented agent builds. |
| [All articles](https://nosistech.com/blog/) | Every NosisTech article, newest first. |
